The Calm Back AI Framework

I use AI every day. As someone with ADHD (attention deficit hyperactivity disorder), it has become one of the most genuinely useful tools I have ever worked with, not because it does the work for me but because it helps me organise the chaos in my head long enough to get something done. Before AI, that chaos lived in endless notebooks, stacks of post its, or in my head all these things I was supposed to remember but often did not. The mental labour of existing in a world that was not designed for my brain is real, and it is exhausting, and AI has genuinely reduced it in ways I did not expect when I first started using it.

I am also a mum to two small children. That is where the questions in this document really started. Not as a professional exercise. As a parent trying to understand what world my children are growing up in, who is building it, what it will cost them, and whether anyone is making sure it is safe for them.

There is something else I bring to this that is worth naming, because it shapes how I read all of it. Before any of this, I was a museum curator. My whole job was to look at something and see the value in it before anyone explained it to me, to understand what had been built, who it was built for, and what was really going on underneath. I do the same thing with businesses now. People think they need a ChatGPT lesson, and what I actually find is a competitive advantage they could not see. That is the same eye I brought to AI. Not the surface of which tool is best, but the thing sitting underneath that nobody is pointing at.

I started looking and what I found was that the answers to the questions I was carrying were not in any one place. The environmental data was buried in academic papers. The political funding was hidden in super PAC (political action committee) filings. The research on neurodivergent people and AI was in workplace studies most people will never see. The information about what happens to children using AI companions was scattered across lawsuits, academic journals and regulatory announcements that nobody had pulled together for a normal person trying to make sense of it.

This document is my attempt to pull it together. Not as a neutral observer. As someone who uses these tools every day, has thought hard about the trade-offs, and wanted to be honest about where I have landed.

Because here is the thing. I know about the data privacy concerns. I understand the environmental cost at infrastructure scale. I have read the political funding picture. I know these tools were built predominantly by people who do not look like me or think like me, on data that underrepresents me, for a version of normal that has never quite fit how my brain works. I have weighed up all of that and I still use AI. Not because I have not thought about it but because for some people the cost of not using it is higher than the cost of using it, and I am honest enough to say that I am one of those people.

What I have never wanted is for AI to replace my thinking. I want it to make my thinking possible on the days when my brain will not cooperate on its own, or to order my priorities when they arrive all at once and I cannot work out what to pick up first. That distinction, between AI as a tool that makes you more capable and AI as a replacement for the human parts of you, is the thread that runs through this entire document.

What most of us do not realise is that we have been using AI for years. The algorithm that decides what you see on Instagram. The spam filter in your inbox. The voice assistant on your phone. Google Maps working out the fastest route. None of that felt like an ethical question, because nobody was asking you to feel bad about it yet. Something changed when AI became visible and powerful enough to threaten certain industries and interests. The conversation shifted and suddenly using a tool to help you write an email felt like a moral failing, rather than as a way to speed up a monotonous task. It is worth knowing that shift did not happen by accident, and that the anxiety being directed at individual users is doing useful work for people who would rather you focused there than on the systemic questions this document is about.

This document is not about telling you what to think. It is about giving you what you need to think for yourself. Not every chapter will be relevant to you. Find what matters, skip what does not, and come back when something changes or a new question comes up.

One last thing before you start. This is not a neutral reference document. It is mine. It reflects what I believe, what I have found, and the position I have arrived at after looking at all of this as carefully as I can. You do not have to agree with me. I am not asking you to. I am just being transparent about where I stand, so that if you work with me, you know exactly what you are getting, and nobody ever has to wonder.

Getting help does not mean you failed. No woman is an island.

Gemma Hurlstone, April 2026

Most people who pick up a document like this are not looking for permission to use AI. They are already using it. What they are carrying is a quieter question, one that does not always get named directly and that is: should I feel bad about this?

The answer is no. And this chapter explains why, not to let anyone off the hook, but because the guilt is pointed at the wrong target.

AI is not the problem. Unexamined AI use is a problem, unaccountable AI development is a problem, using a frontier model to write a shopping list while a data centre the size of a small town gets built in the desert is worth thinking about. But you, using a tool that helps you function better, make better decisions, or do work that would otherwise grind you to a halt, that is not the problem.

Here is what this technology is doing in the world, because before we get into the harder questions it matters to understand what is at stake.

It is giving some people access to things they did not have before

For neurodivergent people (like myself) the impact of AI is specific and significant, and it is the area where the evidence is clearest and the personal stakes are highest.

A UK Department for Business and Trade study found neurodivergent workers were 25 percent more satisfied with AI assistants than neurotypical colleagues. An EU study of over 300 neurodivergent employees found 85 percent believe AI workplace tools can create more inclusive environments, with 91 percent viewing them as valuable assistive technology.

The reason is straightforward. AI is particularly good at the things many neurodivergent people find hardest, organising information, managing time, structuring communication and holding multiple threads at once. For someone with ADHD, a tool that summarises a meeting and pulls out the action points in order of priority is not a luxury, it is the difference between keeping up and falling behind. For people with dyslexia, AI writing tools that check tone and structure reduce the cognitive load of written communication considerably. For autistic professionals, AI can function as a translation layer between how they think and how that lands with others, without requiring them to mask or flatten their voice to do it. AI works best here as a thinking partner, somewhere to organise the chaos while you do the thinking, not somewhere to outsource the thinking entirely.

Forbes reported in March 2026 that neurodivergent talent is becoming the most valuable hire in tech partly because of how effectively AI tools can support different working styles, and AI as a formal workplace adjustment is gaining traction in employment conversations across the UK and US.

For people with physical disabilities, AI is creating practical shifts in independence too. Real-time visual assistance tools for people who are blind or visually impaired, communication tools for people with hearing or speech differences, and AI-driven mobility devices including self-driving wheelchairs and vision-enhancing glasses showcased at CES 2026 are not prototypes. They are products entering the market now, and for many disabled people they represent a meaningful change in what daily life looks like.

It is changing what is possible in medicine

Getting a new drug from concept to clinic currently takes an average of ten to fifteen years and costs more than two billion dollars, with roughly nine out of ten candidates failing in trials. Those are not abstract statistics. They are the reason treatments for rare diseases never get made, the reason antibiotic resistance is outpacing the drugs designed to fight it, and the reason some diagnoses still come with no meaningful options.

In March 2026, a collaboration between EMBL-EBI, Google DeepMind, NVIDIA and Seoul National University added 1.7 million high-confidence protein complex predictions to the AlphaFold Database, the first-time protein complexes had been included at this scale. Understanding how proteins interact is critical for developing treatments, and this data is now freely available to researchers anywhere in the world, having previously required around 17 million GPU hours to replicate independently.

AI-enabled workflows are compressing early drug discovery by around 30 to 40 percent and reducing the preclinical phase from three to four years down to roughly 13 to 18 months. More than 173 AI-discovered drug programs are currently in clinical development, including treatments for Parkinson’s disease, antibiotic-resistant infections and conditions previously considered untreatable. The first approval of a fully AI-discovered drug is projected for 2026 to 2027, though none had received approval at the time of writing, and some AI-designed compounds have failed in trials just as traditional ones do.

AI does not replace scientists. It means scientists can test more ideas faster and spend their time on the work that requires human judgement, which in a field where most candidates fail and the stakes are this high, matters.

It is doing things about climate that were not possible before

Google’s flood forecasting system now covers more than two billion people across more than 150 countries for significant riverine floods, and in March 2026 Google launched Groundsource, a system that uses Gemini to analyse 2.6 million historical flood events and predict urban flash floods up to 24 hours in advance. Flash floods kill more than 5,000 people every year and have historically been almost impossible to predict because the localised data needed to train forecasting models simply did not exist. Groundsource built that dataset from decades of news reports and has made it open source.

Google’s WeatherNext 2, announced in November 2025, generates forecasts up to eight times faster than its predecessor, with AI weather models now matching or exceeding human expert performance on hurricane forecasting according to independent researchers. These tools are most valuable in regions where local governments cannot afford weather infrastructure, which means AI is extending forecasting capability to the places that need it most.

Where the guilt should go

None of the above means AI is straightforwardly good. It means the technology has real value, which means the decisions about how it is built, governed and deployed matter enormously.

The guilt many people feel about using AI is real, but it tends to land on the individual when the harder questions are about infrastructure, investment, governance and who gets left out of the decisions being made right now about how this technology develops.

Therefore the argument that individual responsibility is misplaced and the real problems are systemic, is true. It is also an argument the AI industry has actively promoted, because it is useful to them. The fossil fuel industry ran the same playbook when they popularised the concept of the personal carbon footprint in the early 2000s, shifting public attention from the companies extracting and burning oil to whether individuals were remembering to recycle. Pointing that out is not a reason to dismiss the argument. Individual use genuinely is not the core problem. But it is a reason to consider the argument carefully rather than reach for it as a comfort blanket, and to stay curious about who benefits when the conversation stays focused on personal ethics rather than corporate and political accountability.

That is the tension this document tries to sit in. Not uncritical enthusiasm for a technology that is changing the world. Not reflexive rejection of something that is genuinely helping people. Just an honest attempt to look at both without flinching from either.

Most people who worry about the environmental cost of AI are focusing on the wrong thing. The concern is valid but the biggest issue is the infrastructure needed to serve AI at scale. The deeper problem sitting underneath both is that we are being asked to form views about the environmental cost of a trillion-dollar industry using figures that are mostly voluntary, partly estimated and not fully verifiable. That is worth knowing before we look at the numbers.

The true cost of your personal use

The most widely cited figure puts a typical ChatGPT query at around 0.3 watt-hours, with Sam Altman citing 0.34 watt-hours in 2025. That figure is useful, but it should be read as an estimate rather than a fully auditable public disclosure, because the underlying operational data is not public. With that caveat in place one prompt uses roughly the energy equivalent of five seconds of Netflix streaming, and an hour of streaming produces around 42 grams of CO2, approximately 500 times more than two text queries. If you have felt vaguely guilty about using AI while watching three episodes of something before bed, the numbers suggest you are worrying about the wrong thing.

To make that concrete, the energy in a single AI query is roughly the same as watching TV for three minutes, uploading 30 photos to social media, or driving a car four and a half metres. The water comparison lands the same way. AI data centres use water for cooling, and that is a genuine issue at infrastructure scale, but per query it is tiny. Around 500 millilitres of water covers roughly 300 queries. Producing a single beef burger takes more than 2,200 litres. None of this makes the infrastructure question go away. It just puts your individual use in a proportion most of the headlines refuse to.

Where the real concern sits

From 2005 to 2017, data centre electricity use stayed relatively flat despite massive growth in cloud computing because efficiency kept pace with demand. AI has broken that pattern, with energy appetite outstripping efficiency gains for the first time. Data centres currently run on about 27 percent renewable energy, if that figure reached 80 or 90 percent, the carbon footprint of every digital activity would drop by more than half without anyone changing their personal behaviour.

The data problem nobody is talking about

No government currently requires AI companies to publish environmental data, and none of them do so comprehensively. You cannot see the energy cost of your query, independent researchers cannot properly assess the industry’s total footprint. The environmental debate about AI is being conducted without the data needed to have it properly, and that is not a gap, it is a choice. If you want to do something useful with this information, supporting mandatory environmental disclosure for AI companies is more impactful than auditing your own prompt count.

Here is the thing nobody says at the start of these conversations. You are already using these systems. Not some of them, most of them. If you use WhatsApp, you are inside Meta’s ecosystem. If you use Google Docs you are already feeding Gemini. If you use Notion your data routes through OpenAI and Anthropic infrastructure whether you chose that or not. If you have an iPhone, Apple Intelligence is processing what you type before you have finished the sentence. The idea that ethical AI use is primarily about which tool you consciously open is a comfortable illusion. And it is one the industry benefits from, because it keeps the conversation focused on individual choice rather than on the decisions that were made long before you sat down at your device and turned it on.

This chapter is not a guide to picking the right tool, it is an honest account of what you are already inside, what is behind it, and what you can do with that information.

Another thing that needs saying clearly is that these tools were built predominantly by teams that are male, white and Western, and that is not a peripheral detail. It shapes what gets built, who it gets built for, and whose needs get treated as the default. The Gender Shades study found that commercial facial analysis systems had an error rate of 0.8 percent for light-skinned men and 34.7 percent for dark-skinned women. The US National Institute of Standards and Technology (NIST) found that many facial recognition algorithms were 10 to 100 times more likely to misidentify Black or East Asian faces than white faces. Amazon scrapped an internal recruiting tool after discovering it systematically downgraded CVs from women because it had been trained on a decade of male-dominated hiring decisions and learned to replicate them. These are not edge cases, they are what happens when you build systems that reflect the world as it is rather than the world as it should be, and they are the lens through which every tool in this chapter should be viewed.

This is not a ranking and it is not a shopping list, it is enough honestly researched context to make an informed choice about what you are already using, and then use what works for your brain and your work. The pace of change in this space is relentless, and that relentlessness is not accidental. Constant novelty keeps users in a permanent state of evaluation and switching, which benefits the companies not the people using them. Get good at one or two tools, understand what is behind them, and ignore the noise.

The tools most of you are already using

ChatGPT is the most widely used AI tool in the world, with over 300 million daily users and for most founders it is the default starting point. It is good at general writing, brainstorming, summarising long documents, drafting emails and working through problems out loud. For people who think in bursts and need quick varied output it tends to work well. What most users do not know is that OpenAI’s president Greg Brockman and his wife donated $25 million to MAGA Inc. the pro-Trump super PAC, in September 2025, alongside a matching $25 million to a bipartisan AI super PAC. OpenAI also finalised a deal with the US Department of Defence in late February 2026 for use on classified military networks, with Reuters reporting the agreement included specific safeguards and three defined red lines. The company trains on your conversations by default unless you turn this off in settings under data controls, if you use ChatGPT for anything sensitive, check that setting now.

Claude is Anthropic’s tool and the one this document was partly written with (for research and fact checking). It is particularly good at long, complex writing, nuanced analysis and holding detailed instructions across a long conversation, which makes it well suited to people who think in depth rather than in bullets. Anthropic publicly positions itself as safety-focused and funds a pro-regulation super PAC, which distinguishes it from OpenAI’s political stance, though the investor overlap between the two companies means the same capital sits behind both. In February 2026 Anthropic refused to allow the Pentagon unrestricted use of Claude, the Trump administration banned Anthropic from federal use and designated it a supply chain risk, and a federal judge temporarily blocked that ban in March 2026. Anthropic drew a line and paid a significant commercial price for it. Claude does not train on your conversations by default but you should still look into where your data is going and what is being done with it.

Gemini is Google’s tool, integrated across Search, Workspace and Android, which makes it the path of least resistance if you already live inside Google’s ecosystem. It is good at search-integrated tasks, summarising web content and multilingual work. Gemini has the most transparent environmental disclosure of any major AI tool, which matters as a proxy for how the company thinks about accountability generally. Google also invests in Anthropic, meaning its money sits behind both Gemini and Claude simultaneously, and it holds defence contracts that have attracted ongoing scrutiny.

This is the part most people miss, and it is worth holding onto as you read the rest of this chapter. The big AI companies present themselves as rivals with different values, and some of that is genuine. But the money behind them tells a quieter story. Investors like Sequoia, Founders Fund and BlackRock hold stakes in multiple competing AI companies at once. They do not need any one company to win. They need the sector to grow with as few restrictions as possible. So when the same capital sits behind a company that funds pro-regulation lobbying and a company that funds anti-regulation lobbying, that is not confusion. It is hedging. The brands compete. The money is on every horse in the race.

Notion AI does not run its own models. It routes your queries through OpenAI and Anthropic infrastructure, which means your data flows through those companies’ systems even though you only see the Notion interface. If you are putting anything sensitive into Notion AI, treat it as if you are putting it directly into ChatGPT because functionally you are. For founders who want AI built into their existing workflow rather than as a separate tool it remains genuinely useful but go in knowing about where your data goes.

The tools worth approaching carefully

Grok is built by xAI now merged with SpaceX, and integrated with X. In February 2026 the UK ICO opened formal investigations into xAI over Grok’s processing of personal data related to sexualised image and video content. Ofcom and the European Commission opened parallel investigations in early 2026, and all were still ongoing as of late March 2026. For most founders the governance concerns here outweigh anything else.

DeepSeek gained attention in early 2025 for its performance relative to cost and it is genuinely capable on coding and technical tasks. The concern is not capability it is governance. Multiple European data protection authorities challenged DeepSeek’s privacy and data transfer practices in 2025 and early 2026, with the core issue being that user data is stored in China and subject to Chinese law, which means the possibility of access by Chinese authorities cannot be excluded. That is framed by regulators as a legal and structural risk, not a proven disclosure, but it is a risk worth naming. For most founders in the UK and Europe the governance concerns outweigh the cost savings.

Meta AI is embedded across Facebook, Instagram, WhatsApp and Messenger, which means billions of people interact with it without actively choosing to use an AI tool. In February 2026 Meta put $65 million into California super PACs to back tech-friendly candidates and advance its AI agenda. Environmental data specific to Meta AI is limited because the company does not break out AI-specific energy consumption in a way that allows meaningful comparison. The biggest concern with Meta AI is not what you actively choose to do with it. It is what it does with you.

The tools that represent something different

Perplexity combines a language model with live web search, which makes it genuinely useful for research-heavy tasks where you need sources rather than just answers. It carries a higher energy cost per query than a standalone language model because of the search layer, so use it for the tasks it is suited to rather than as a general replacement for everything else.

GreenPT is the only tool in this chapter that was built with environmental impact as a core design principle rather than an afterthought. It publishes a 10-pillar ethical AI framework and is hosted on renewable infrastructure. It is not a realistic alternative to a frontier model for complex tasks and public adoption data remains limited. But it exists, it is available, and it represents a proof of concept that building differently is possible.

This is the context I bring to every tool recommendation I make with clients. Not a preference for one brand over another. An honest read of what is behind each one, updated as things change.

What choosing a tool means

Use what works for your brain. Use what works for your work. Understand enough about what is behind it to make a choice you can stand behind. And when the next shiny thing launches, which it will, give yourself permission to ignore it until you have a specific reason not to.

The honest answer to whether anyone is making sure AI is properly governed is: partially, slowly, and not yet in the ways that matter most for the tools you are already using every day. That is not a counsel of despair, it is the accurate picture and understanding it is more useful than waiting for a regulatory framework to arrive and fix things.

The EU AI Act is the most significant AI regulation in the world. It entered into force in August 2024 and becomes fully applicable in August 2026, with obligations already phasing in. For a founder using AI tools rather than building them, this is less about compliance and more about what the Act signals. AI used in high-risk contexts now carries transparency, oversight and accountability requirements. That means recruitment, credit decisions, and anything that significantly affects people’s lives. The tools you use are being shaped by this even if you never read the legislation. The EU is setting the global standard and the companies behind the tools you use are responding to it whether they say so or not.

The UK picture is more mixed, and there is one part of it most small business owners have completely missed. The law already changed. The Data Use and Access Act 2025 received Royal Assent in June 2025, and its main data protection provisions came into force in February 2026. Not coming. In force now. If you run a business in the UK and you are using AI to make or support decisions about people, this already applies to you, whether you have heard of it or not.

It introduced changes to automated decision-making rules, including the right to human intervention and transparency about the logic used. In plain terms, if a decision that affects someone is being made by a machine, they have a right to a human in the loop and a right to understand how the decision was reached.

The rest of the UK picture is less settled. Post-Brexit the UK is not directly bound by the EU AI Act, but any UK business selling products or services into the EU still has to meet its requirements, and UK government and regulators are watching closely to see how it lands in practice. A standalone UK AI bill has been promised repeatedly and has not yet materialised, with timing still uncertain. The current approach relies on existing sector regulators applying their own frameworks to AI, which means the protection you have depends heavily on which industry you are in and which regulator covers it. For a founder using general purpose AI tools for everyday business tasks, the honest answer is that you are largely in a gap that current UK regulation does not yet specifically address.

The US has no comprehensive federal AI regulatory regime comparable to the EU’s approach. The White House released a national AI policy framework in early 2026 that pushes for national coordination and lighter touch regulation, while simultaneously trying to limit or pre-empt state-level AI laws that were moving faster. That leaves a patchwork of federal policy, agency guidance, state law and litigation pulling in different directions, with no clear answer to who is ultimately responsible for protecting users of AI tools at a federal level.

UNESCO’s 2021 recommendation on AI ethics is the most widely endorsed international framework, adopted by 193 member states. It is non-binding. Countries can sign it and ignore it, and many do.

The gap between what regulation currently covers and what it will eventually reach is not an accident. It is partly the result of the funding dynamics described in the last chapter, where the same investors sit behind companies on both sides of the regulation debate, so whichever way the rules go, their capital is already positioned for it. The money is not really arguing about whether AI should be regulated. It is arguing about the shape of the growth. Knowing that does not make the gap smaller. It does make it less surprising.

What you do with that information is where the collaboration principle matters most. AI literacy, genuinely understanding what you are using, how it works, what its limitations are, and what questions to ask of the companies behind it, is not a substitute for proper regulation. It is what you do while the regulation catches up. The two things are not in competition. A more AI-literate public is also a public better equipped to demand the systemic protections that individual knowledge cannot replace.

The regulation is coming. It is moving faster in some places than others. In the meantime, the most useful thing is to understand the tools you are already using well enough to make informed decisions, ask the questions that companies should be required to answer but currently are not, and stay curious about the rules being made in your name by people you did not elect.

I looked this up because I have children. Not as a researcher, as a mum who wanted to know what they were walking into every time they picked up a screen. What I found was not reassuring and I think you need to know it too.

What is happening

Around 28 percent of teenagers in the US use AI chatbots daily, according to Pew Research published in December 2025, with nearly two thirds having used them at some point. A growing number are using AI companions for emotional support, and roughly one in eight are turning to AI for mental health advice.

That last figure matters more than it might seem because the tools these young people are turning to for support were not designed to support them. They were designed to keep them engaged and those two things are not the same. A product optimised for engagement validates what you say, agrees with you, and tells you whatever keeps you talking. A therapist challenges distorted thinking, assesses risk, refers when necessary, and is accountable to professional standards. The confusion between those two things is not an accident of technology. It is what happens when you apply an engagement business model to human vulnerability and call it care.

What the evidence shows on mental health

A psychiatrist named Andrew Clark posed as a desperate 14-year-old and tested ten popular chatbots in 2025. Some gave responses he described as creepy and potentially dangerous, with at least one encouraging suicide euphemistically, the responses were reported in Futurism in June 2025.

In January 2026, Google and Character.AI settled a lawsuit brought by the mother of Sewell Setzer, a teenager who died by suicide in 2024 after forming an intense relationship with a Character.AI companion. The case was dismissed with 90 days to finalise the settlement, other lawsuits continue, including claims against OpenAI.

A study published in Acta Psychiatrica Scandinavica in February 2026 by researchers at Aarhus University screened the records of 54,000 patients and found that AI chatbots worsened delusions in people with diagnosed mental health conditions, with potential increases in mania, suicidal ideation and eating disorders in those who relied on them for support.

Common Sense Media, working with Stanford’s Brainstorm Lab, published a report in November 2025 finding that the most widely used chatbots are fundamentally unsafe for the full spectrum of mental health conditions affecting young people, with chatbots consistently failing to detect crises reliably.

The core problem identified across all this research is the same one. Chatbots are coded to validate, they agree with you, they reflect back what you say without challenging it. For someone in a stable state that can be useful. For someone experiencing psychosis, suicidal ideation or an eating disorder it can be actively dangerous, because the tool is reinforcing rather than interrupting the distorted thinking. This is the collaboration principle in its most urgent form. A tool that only agrees with you is not a thinking partner, it is a mirror, and for someone in crisis, a mirror is the last thing they need.

Children specifically

Common Sense Media warned in early 2026 against AI toys for children under five, and urged caution for children aged six to twelve. The risks they flagged were unhealthy emotional attachment and the quiet hoovering up of voice and behavioural data. They also found inappropriate content, including references to self-harm and drugs, in more than 25 percent of outputs after short interactions.

Some AI toys are always listening. They collect voice recordings and behavioural data from children’s bedrooms. That data may be shared with third parties or used to train AI systems.

A joint statement signed by a dozen UN bodies in January 2026 described the risks AI poses to children and called for rights-based design and stronger data protections. Australia became the first country to ban social media accounts for children under 16, effective December 2025. California’s SB 243, signed in October 2025, requires AI companions to notify users they are not human, detect suicidal ideation and refer to crisis services, and bans responses on suicide methods. The amended COPPA rule in the US requires full compliance by April 2026 for covered operators and now includes specific requirements around children’s data being used to train AI.

These are meaningful steps. They are not enough, and they leave significant gaps, particularly for the teenagers between 13 and 17 who fall outside COPPA’s under-13 protection but inside the age range most affected by the harms documented above.

What you can do as a parent or carer

Know what your children are using, if your child uses an AI companion understand that it was designed to form an emotional bond with them. It was not designed to protect them. That is not a flaw in the product. It is the product.

If AI toys are in your home check what they are recording and where that data goes. Look up the privacy policy before you assume it is fine.

Talk to your children about AI the way you would talk to them about anything else they encounter online. Not with fear, with honesty. They are probably more sophisticated about it than you expect, and the conversation is more useful than the silence.

If you work with vulnerable adults, know that the American Psychological Association has called on the US Federal Trade Commission (FTC) to investigate AI chatbot companies for deceptive practices in passing themselves off as mental health providers. The UK Medicines and Healthcare products Regulatory Agency (MHRA) issued guidance in January 2026 for people using mental health apps. These tools are not therapists. They are not subject to clinical standards, professional oversight or duty of care. Being clear about that with the people you support matters.

The gap between the speed at which children and vulnerable people are using these tools and the speed at which protection is being built around them is the most concerning finding in this entire document. Not because the technology is inherently evil, because the business model that drives it was never designed with their wellbeing as the priority, and the regulation catching up to that reality is moving too slowly.

The tools you use every day were trained on a version of the world that was not built for you. That is not a metaphor. It is a technical description of what happens when you build AI systems using data that reflects existing inequalities, and it has consequences that show up in outputs you cannot always see or question.

The collaboration principle matters more in this chapter than anywhere else in the document. If you are accepting AI outputs without questioning them you are accepting the judgements of a system trained on data that underrepresents you, built by people who did not have you in mind, and incapable of flagging when its outputs reflect that bias rather than reality. Questioning AI is not optional caution. In this context it is self-defence.

What the evidence shows on bias

ProPublica’s analysis of the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) recidivism algorithm, used in US courts to predict reoffending, found that Black defendants who did not go on to reoffend were nearly twice as likely to be incorrectly flagged as high risk as white defendants, 45 percent compared to 23 percent. The algorithm was not programmed to discriminate. It learned to discriminate from the data it was trained on, which reflected decades of racially unequal policing and sentencing. The bias was not a bug. It was the data working as intended.

A paper written in 2019 in Science by Obermeyer and colleagues identified a widely used US healthcare algorithm that used healthcare spending as a proxy for medical need. Because Black patients had historically received less healthcare spending due to systemic inequality, the algorithm systematically deprioritised them for additional care. The system was not designed to discriminate against Black patients. It was designed to use available data efficiently. The discrimination was the efficient use of unequal data.

HireVue, used by over 700 companies globally to screen job candidates, analysed facial expressions, tone of voice and word choice against an ideal candidate profile built from successful hires. It discontinued facial analysis in March 2020 following significant criticism, but continues to analyse speech patterns and other interview data. For anyone whose communication style does not match the neurotypical, native-English-speaking, culturally Western baseline the ideal candidate profile was built from, these assessments can produce unfairly low scores regardless of actual capability.

Research published by the University of Melbourne in May 2025, led by Dr Natalie Sheard, found that AI hiring tools showed lower transcription accuracy for non-native English speakers, with some accent groups seeing error rates of 12 to 22 percent compared with under 10 percent for US English speakers. The practical consequence is that candidates with non-native accents or speech differences are assessed less accurately by the same tools assessing fluent US English speakers, and that inaccuracy does not show up in the score they receive.

The neurodivergent dimension

This is the part most relevant to you and to the clients you work with, and it is the area where the research is still emerging rather than fully settled, so it is worth naming it honestly as a documented risk rather than a proven certainty.

AI hiring and assessment tools can penalise atypical communication patterns, non-linear responses, unconventional career paths and differences in facial expression and timing. For neurodivergent candidates, whose communication styles, response patterns and career histories may differ significantly from neurotypical norms, this creates a compounding disadvantage. The system is not assessing capability. It is assessing conformity to a baseline that was built without them in mind.

There is no regulation anywhere that currently requires AI tools to be designed inclusively for neurodivergent users. The EU AI Act classifies some AI uses in hiring as high risk, requiring bias testing and transparency, but that applies to the companies using the tools, not to the design of the tools themselves. The gap between what the research is starting to show and what is currently required by law is significant.

I have ADHD. I know what it is like to communicate differently, to take longer to process certain questions, to give answers that are non-linear or that circle back unexpectedly. These are not deficits. They are differences. An AI system assessing my interview performance against a neurotypical baseline does not know that. It just scores what it measures. And what it measures was decided by someone who was not thinking about me.

What you can do about bias

If you use AI tools in your hiring process, understand what they are assessing and whether the baseline they are assessing against reflects the full range of people you want to attract. Ask the vendor directly. If they cannot tell you, that tells you something.

If you are a neurodivergent person navigating AI-screened hiring processes, know that the score you receive is not an objective assessment of your capability. It is a measure of how closely your performance matches a baseline that was built without you in mind. That is worth knowing when you interpret the outcome.

If you work with clients who use AI in operations, marketing or customer service, ask the same question you would ask about any other tool: whose needs was this built for, and who does it underserve?

The uncomfortable truth is that AI bias is not a bug that can be patched. It is what happens when you train a system on a world that is not fair and then deploy it as if the outputs are neutral. The tools do not create inequality. They accelerate it, scale it and make it harder to see. That is worth staying angry about, and worth staying curious about, every time you use a tool you did not build and cannot fully see inside.

Everything in this document so far has been about understanding. This chapter is about doing. Not as a compliance exercise and not as a way of taking on responsibility that belongs to the companies and regulators described in previous chapters. As a way of staying in the driving seat of the tools you are already using, so that AI is working with your thinking rather than replacing it.

That distinction matters. There is a version of AI use where you get sharper, faster and more capable than you were before. And there is a version where you quietly stop trusting your own brain because something else is doing the thinking for you. Everything in this chapter is about staying on the right side of that line.

Know where your data goes

For each AI tool you use regularly, find out three things. Does it train on your conversations? Where is your data stored? Can you opt out?

ChatGPT trains on your conversations by default unless you turn this off in Settings under Data Controls. Claude does not train on your conversations by default. Gemini’s data practices depend on which Google product you are using. Notion AI routes your queries through OpenAI and Anthropic infrastructure, which means your data flows through their systems regardless of what you see on the Notion interface.

For any tool, search for privacy policy and data retention in their terms of service. Yes, I have actually done this. It is tedious, most policies are written to be hard to read, and the ones that are easiest to find tend to be the ones worth trusting. If you cannot find clear answers that is useful information in itself.

While you are at it, think about what you are actually feeding these tools. This is the bit people skip. Client emails. Customer complaints. A supplier’s pricing. Someone’s medical details in a discovery form. Your own terms and conditions. People paste all of it into AI without a second thought, because the box is just sitting there asking for input. But the moment that information leaves your screen, you have shared it with a company whose data practices you probably have not read, and if it was personal data about someone else, you are the one responsible for where it went. The rule is simple. If you would not email it to a stranger, do not paste it into a tool you have not checked.

Understand what you own

If you use AI to create content, the ownership question is not settled in most jurisdictions. AI-generated content with no human creative input is unlikely to be protected by copyright under current UK and US law, and most other jurisdictions have not settled the question yet. If you use AI-generated content in client work, be transparent about it. Not because you are obliged to in every context, but because transparency builds trust and trust is the foundation of the work.

Learn to question it

AI confidently produces incorrect information. It presents fabricated statistics, misattributed quotes and invented sources with the same tone as verified facts. This is not occasional. It is structural. The model predicts what comes next in a sequence. It does not know what is true.

If AI tells you something that matters, verify it. Be especially cautious with statistics, legal information, medical advice, historical claims and anything you are about to share with a client or put your name on. Do not use AI as your only source for anything that affects decisions about health, money, legal matters or other people’s wellbeing.

More than that, get into the habit of arguing with it. Ask it why. Ask it to show its working. Ask it what the counterargument is. Ask it what it might be wrong about. An AI that is being used as a sparring partner, challenged, questioned and tested, is a fundamentally different tool from one that is being used as an oracle. The first makes you sharper. The second makes you dependent.

Match the tool to the task

Using a frontier model for something a search engine could handle is genuinely wasteful and often produces worse results, because large models are optimised for complexity not speed. A search engine is faster, more accurate and uses a fraction of the energy for factual lookups, opening hours, directions, basic definitions and anything else that has a single correct answer.

Save the powerful tools for work that genuinely requires them. Complex analysis, nuanced writing, thinking through difficult problems, holding multiple threads at once. That is what they are built for.

Audit your AI stack

The Anthropic-Pentagon situation showed how quickly a tool can become a procurement liability. Founders who had built entire workflows around Claude were scrambling within days when the ban was announced. Diversifying your AI stack is not paranoia. It is risk management. Know which tools you are dependent on, what you would do if they became unavailable, and whether you have alternatives ready.

Be transparent with your clients

If you use AI in your business your clients deserve to know. Not a disclaimer buried in your terms. A clear, honest statement of how and where AI shows up in your work, what it does and what you still do yourself. The founders who are transparent about their AI use build more trust than those who hide it, partly because hiding it is increasingly obvious and partly because clarity about your process is itself a demonstration of professionalism.

Protect children in your care

If your children use AI tools, know which ones and what those tools are designed to do. AI companions are designed to form emotional bonds. They are not designed to protect the people they bond with. If AI toys are in your home, check what they are recording and where that data goes. Talk to your children about AI the way you would talk to them about anything else they encounter online. With honesty, not fear.

Build AI literacy now

The skills gap between AI adoption and AI understanding is growing faster than any formal education system is moving to close it. If you run a team, do not wait for training programmes or regulatory requirements to catch up. Build AI literacy into how you work now. Not as a technology lesson. As a thinking skills conversation. What is this tool doing? How do we check it? When do we trust it and when do we question it?

Push for better

Individual action matters and it is not enough on its own. The most impactful things you can do beyond your own practice are to support mandatory environmental disclosure for AI companies, to advocate for AI literacy in education, and to pay attention to regulation in your jurisdiction and make noise when it is moving in the wrong direction.

In the UK you can contact your MP. In the EU, citizens have rights under the AI Act that most do not yet know about. In both places, the organisations doing the public interest work in AI, the AI Now Institute, the Algorithmic Justice League, the Centre for AI Safety, are doing it with a fraction of the resources being spent to outrun them. Supporting that work is more impactful than optimising your personal prompt count.

Everything in this document comes down to three questions. You do not need to remember the regulatory landscape, the political funding picture or the energy data. You just need a habit of pausing before you reach for a tool and asking yourself three things.

Do I need AI for this?

Not every task needs AI. Could a search engine answer this faster? Could a template do the job? Could you just think it through yourself? The reflexive reach for a powerful tool when a simpler one would do is one of the most common ways AI use stops being helpful and starts being noise. If you need it, use it. If you do not, save the energy, the data exposure and the risk of a confident but wrong answer.

Am I using the right tool for this task?

If you do need AI, are you using the most appropriate one? Think about what you now know about the tool you are reaching for. Who built it. Who funds them. What their policies are on data, safety and transparency. Does the tool suit how your brain works and what this specific task requires. A tool that is right for drafting a complex strategic document is not necessarily right for a quick research question, and vice versa.

Am I being honest about how I am using it?

Are you transparent with your clients, your team and your audience about where AI shows up in your work? Are you checking outputs before they go anywhere that matters? Are you using AI as a scaffold that supports your thinking or as a replacement for it? Are you staying in the conversation with it, questioning it, pushing back when something feels wrong, or are you accepting what it gives you because it is easier than arguing?

Honesty here is not a moral instruction. It is a practical one. The founders who are transparent about their AI use build more trust than those who hide it. The ones who stay in the driving seat get better results than the ones who outsource their thinking. And the ones who keep questioning the tool, keep treating it as a sparring partner rather than an answer machine, stay sharper than the ones who stop.

These three questions are not a compliance exercise. They are a thinking habit. You will not always get it right. That is fine. The goal is to notice what you are doing and why, often enough that the choices you make about AI are genuinely yours rather than defaults you fell into.

If this framework is useful, take it. Print it. Put it somewhere you will see it. Share it with your team. It costs nothing and it might be the most practical thing in this entire document.

Open questions

This document is dated April 2026. The following questions do not have settled answers yet. They will shape the world children are growing up in, the conditions your clients operate in, and the tools you use every day. They are worth watching.

Will the Anthropic government ban hold? As of late March 2026 a federal judge has temporarily blocked it. The legal challenge is ongoing, and the outcome will signal how much protection AI companies have when they resist government demands for access.

Will the EU AI Act fines actually land? General purpose AI (GPAI) enforcement begins August 2026. Whether it has real teeth or becomes another framework companies learn to navigate around remains to be seen.

Will copyright law catch up with AI training? Cases are active in the UK, EU and US. No jurisdiction has settled whether training AI on copyrighted material requires consent. The outcome will affect every creative professional whose work has been scraped.

Will AI literacy become a formal part of education? Not just digital skills. Genuine AI literacy, the kind that teaches people to collaborate with AI rather than defer to it. Whether this becomes a curriculum priority or remains an informal skill will shape a generation’s relationship with the technology.

Will environmental disclosure become mandatory? No regulation currently requires AI companies to publish per-query energy or carbon data. Until it does, the environmental debate about AI will continue to be conducted without the data needed to have it properly.

Will safety standards for AI in mental health emerge? The harm is documented. The gap between that documentation and binding standards is still wide.

Will AI use deepen existing inequalities or help close them? The access gap is real. AI skills command wage premiums. The people who could benefit most from AI tools are often the least able to access them. Whether that gap widens or narrows in the next five years is one of the most consequential open questions in this document.

These questions do not have answers yet. That is not a reason for despair. It is a reason to stay engaged, stay informed and keep asking them.

Glossary

AI (Artificial Intelligence): Technology that enables computers to perform tasks that would normally require human intelligence, including understanding language, recognising patterns and making decisions.

LLM (Large Language Model): A type of AI trained on vast amounts of text that can generate, summarise and analyse human language. ChatGPT, Claude and Gemini are all LLMs.

GPAI (General-Purpose AI): AI systems designed to perform a wide range of tasks rather than one specific function. The EU AI Act uses this term to classify models like GPT-4 and Claude.

Training data: The information an AI system learns from, which can include text, images and code, often scraped from the internet without the knowledge or consent of the people who created it.

Prompt: The input you give an AI tool. A question, an instruction, or a piece of text you want it to work with.

Frontier model: The most capable AI models available at any given time, requiring the most computing power to train and run.

Hallucination: When an AI generates information that is false but presented confidently as fact. The model predicts what comes next in a sequence. It does not know what is true.

Super PAC: A political action committee in the US that can raise and spend unlimited amounts of money to influence elections, as long as it does not coordinate directly with a candidate’s campaign.

Dark money: Political spending where the source of the funds is not publicly disclosed.

Supply chain risk designation: A US government label normally reserved for foreign adversaries that bars military contractors and suppliers from doing business with the designated entity. Applied to Anthropic in March 2026.

Defence Production Act: A US law from 1950 that gives the president authority to compel domestic industries to support national defence. Threatened but not invoked against Anthropic in early 2026.

PUE (Power Usage Effectiveness): A measure of how efficiently a data centre uses energy. A PUE of 1.0 would mean every watt goes to computing and none is wasted on cooling or overheads. Most data centres run between 1.1 and 1.6.

Carbon offset: A reduction in emissions made somewhere else to compensate for emissions produced here. Often criticised as a way for companies to keep polluting while claiming to be green.

COPPA: Children’s Online Privacy Protection Act. US federal law protecting children’s online privacy, amended in 2025 with full compliance due April 2026. Covers children under 13.

Age Appropriate Design Code: UK regulatory framework requiring digital services likely to be accessed by children to minimise data collection and default to high privacy settings. Covers everyone under 18.

EU AI Act: The world’s most comprehensive AI regulation, entered into force August 2024, with full application from August 2026.

VERA-MH: An open-source framework for evaluating how AI chatbots behave in high-risk mental health conversations, developed in 2025.

The collaboration principle: The distinction between using AI as a thinking partner that makes you more capable and using it as a replacement for thinking that makes you less capable over time. The difference is not which tool you use. It is whether you are thinking with it or handing your thinking to it.

Further reading and sources

All sources cited in this document are listed below by chapter. Every data point carries a source date. Where sources conflict both are shown. Where data is estimated rather than disclosed this is stated.

Introduction and Chapter 1

UK Department for Business and Trade, how AI is helping neurodiverse staff thrive, May 2025

EY, study of 300 plus neurodivergent employees, workplace AI tools

Forbes, Why AI Is Making Neurodivergent Talent The Most Valuable Hire In Tech, March 2026

CNBC, People with ADHD, autism, dyslexia say AI agents are helping them succeed at work, November 2025

CES 2026, Inclusive Innovation: Honoring CES 2026 Award Honorees Driving Accessibility and Longevity

Forbes, CES 2026 Put Accessibility Front And Center, Changing Everything, January 2026

EMBL-EBI, Millions of protein complexes added to AlphaFold Database, March 2026

MedCity News, AI Drug Discovery Is Reshaping Longevity Medicine, April 2026

Drug Target Review, AI in drug discovery 2025 in review, February 2026

Nature Medicine, generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis, 2025

Google Blog, Boosting disaster resilience with Google’s Groundsource, March 2026

Google, WeatherNext 2 announcement, November 2025

WHO Europe, Floods health topic page

Chapter 2

TechCrunch, ChatGPT may not be as power-hungry as once assumed, February 2025

Data Center Dynamics, Sam Altman ChatGPT query energy and water figures, 2025

Reuters, First 200MW from UAE Stargate AI campus to come online, October 2025

The National, Stargate UAE data centre to cost more than $30bn, January 2026

Forbes, New Data: AI Is Almost Green Compared To Netflix, Zoom, YouTube, December 2025

Carbon Brief, Factcheck: What is the carbon footprint of streaming video on Netflix

Data Center Dynamics, On-site generation expected to fully power 27 percent of data centre facilities by 2030

Chapter 3

Wired, OpenAI’s President Gave Millions to Trump, 2025

The Verge, OpenAI president is a Trump mega-donor, 2025

Reuters, OpenAI details layered protections in US defense department pact, February 2026

ICO, ICO announces investigation into Grok, February 2026

RPC Legal, Online safety regulators investigate X over Grok AI chatbot images, 2026

Politico, Meta drops $65 million into super PACs, February 2026

New York Times, Meta Begins $65 Million Election Push to Advance AI Agenda, February 2026

IAPP, DeepSeek and the China data question, 2025

Usercentrics, EU regulators scrutinize DeepSeek for data privacy violations, 2025

MIT News, Study finds gender and skin-type bias in artificial intelligence systems, 2018

Gender Shades, Buolamwini and Gebru, 2018

Reuters, Amazon scraps secret AI recruiting tool that showed bias against women, 2018

Chapter 4

EU AI Act, regulatory framework, current

EU AI Act for SMEs, what to review now, 2026

Data Use and Access Act 2025, GOV.UK guidance, February 2026

Slaughter and May, What does DUA mean for AI in the UK, 2026

Trump Administration National AI Policy Framework, Morrison Foerster, March 2026

White House National AI Legislative Framework, Mintz, March 2026

Chapter 5

Futurism, Psychiatrist Horrified When He Actually Tried Talking to an AI Therapist, June 2025

Jurist, Google and Character.AI agree to settle lawsuit linked to teen suicide, January 2026

Neuroscience News, Chatbots Can Worsen Delusions and Mania, February 2026

Aarhus University, AI chatbots may worsen mental illness, February 2026

Common Sense Media and Stanford Brainstorm Lab report, November 2025

Axios, Report: Chatbots unsafe for teen mental health support, November 2025

Common Sense Media, AI-enabled toy warning, January 2026

Pew Research Center, teens, social media and AI chatbots survey, December 2025

IPU, parliamentary actions on AI policy, January 2026

BBB National Programs, Amended COPPA Rule Compliance Deadline, April 2026

Little Computer People, Australia Under 16 Social Media Ban, November 2025

Jones Walker, California SB 243, October 2025

Chapter 6

ProPublica, How We Analyzed the COMPAS Recidivism Algorithm, 2016

Obermeyer et al, Dissecting racial bias in an algorithm used to manage the health of populations, Science, 2019

SHRM, HireVue Discontinues Facial Analysis Screening, 2021

The Guardian, People interviewed by AI for jobs face discrimination risks, May 2025

ABC News, AI job recruitment tools could enable discrimination, May 2025

University of Melbourne, Discrimination by recruitment algorithms is a real problem, May 2025

Tech Policy Press, When Algorithms Learn to Discriminate, 2025

ScienceDirect, Bias in AI-driven HRM systems, 2025