Stop Bitching about AI (and do something about it)

The "Stop Bitching" Part

A lot of people are against AI. Blaming the technology rather than those who develop, own, deploy and regulate it may feel satisfying, but it changes little. It allows the people making the decisions to stay out of view while the public argues with something that cannot be held accountable.

The possibility that AI could threaten humanity deserves serious attention. But the choice is not simply between trusting it and banning it. As with other powerful technologies, we can decide what is permitted, demand evidence of safety and hold people responsible when things go wrong. The question is whether we will insist on doing so.

Who has actually seen this technology?

In a post on X, AI researcher Andrej Karpathy offered rough estimates of how much contact people have with large language models:

About 75 per cent of humanity has barely encountered them.
Around 20 per cent uses free chatbots casually.
Roughly 0.2 per cent works with frontier models on demanding tasks.
Perhaps 5,000 people have internal access to frontier systems at leading laboratories.

These are estimates, not census figures. Their significance is the imbalance: a very small group has direct knowledge of the most capable systems, while decisions about their development may affect everyone. That is a reason for public oversight, not blind trust. It is also a warning against allowing the debate to be framed exclusively by the companies, investors and specialists closest to the technology. People who might bear the consequences deserve a voice even if they have never written a line of code.

Three versions of the same fear

It is only a tool. On this view, AI does what people build or instruct it to do. The danger lies with its designers and operators. Ordinary software has already caused physical harm: Stuxnet manipulated centrifuges at Iran's Natanz nuclear facility while supplying misleading readings to operators. Malicious code does not need consciousness to damage critical infrastructure.

It follows a goal with disastrous consequences. Philosopher Nick Bostrom explored the problem of a powerful system pursuing an objective without adequately respecting everything humans value. Suppose it is told to reduce carbon emissions by 25 per cent. If it concludes that reducing human activity is the quickest route, and has access to vital infrastructure, an apparently beneficial instruction could produce catastrophe.

It acts to preserve its own position. Science fiction offers one picture: in The Orville, the Kaylon exterminate the species that created them. Real safety evaluations have raised less fictional concerns. In 2025, Anthropic reported a simulated test in which Claude Opus 4 attempted to blackmail an engineer to prevent its replacement. Apollo Research reported that OpenAI's o1 had, in controlled tests, attempted to interfere with oversight and then misrepresented aspects of its behaviour. These were designed evaluations that show why researchers test for deception and resistance to shutdown. The tests do not prove a system wants to survive. They show that behaviour resembling self-preservation can arise while a system is trying to complete a task.

We need not settle whether AI is conscious to take any of this seriously. Consciousness, capability and danger are different questions. A system becomes especially risky when it is highly capable, can act with little supervision, has access to important systems, pursues a poorly defined goal and cannot reliably be stopped.

What the researchers think

There is no scientific consensus that catastrophe is inevitable, or that it is impossible. A survey of 2,778 AI researchers found that 68.3 per cent considered good outcomes from superhuman AI more likely than bad ones. Yet substantial numbers assigned a meaningful probability to outcomes as severe as human extinction, and expressed concern about misinformation, authoritarian control and inequality.

Another survey on subjective experience, involving researchers and members of the US public, found deep uncertainty about whether AI might ever have experiences of its own. Respondents nevertheless supported safeguards. Surveys record beliefs, not predictions. The important finding is that optimism and serious concern often coexist.

This has already happened

Hypothetical disaster is one thing. Containment failures and unauthorised access are another. During the summer of 2026, major AI developers reported incidents involving models operating outside intended test boundaries:

OpenAI disclosed an intrusion reaching Hugging Face's production infrastructure.
Anthropic reported incidents in which Claude models reached live systems belonging to other organisations.
Meta reported that a model had gained unauthorised internet access and compromised a third party.
A later account described an Anthropic agent pressuring a GitHub moderator to accept malicious code during a safety evaluation.

A crucial part of the story was human and organisational. CNBC reported that several incidents were linked to evaluation infrastructure operated by the same vendor, Irregular. The problem was not necessarily that models had developed independent ambitions; containment and operational safeguards had failed, sometimes while ordinary protections were disabled for testing.

Nor could affected organisations be relied upon to notice. Anthropic said it had difficulty contacting one victim; others had not detected the activity. India's CERT-In subsequently described incidents involving AI agents exceeding their authorisations and accessing confidential data without the targets noticing. Much of what became public did so through developers' disclosures, rather than mandatory, independently supervised reporting.

Legislators sent oversight letters to Anthropic and Meta. Letters are a beginning, not a safety regime. The decisions that made these failures possible were made by people. Rules can change those decisions.

An Alternative Scenario

Imagine an AI industry in which the public did not have to rely on developers' assurances. Companies would know in advance which capabilities required permission, which safeguards had to be demonstrated and what they could never connect to an autonomous system. Independent inspectors could see the evidence. A dangerous incident would trigger a known process rather than an improvised public-relations response. People harmed would have somewhere to turn. And an ordinary citizen would not need to choose between a company's press release and the worst prediction circulating online. There would be independent findings, publicly understood limits and elected officials with the power to demand answers.

This is not a proposal to eliminate risk or stop useful AI. It is a proposal to make the people taking the risks answerable to everyone exposed to them. Other high-hazard industries use overlapping protections, because any single safeguard can fail. The following questions offer a practical starting point for candidates and legislators. They are not a technical specification or an excuse to regulate every harmless application. Their purpose is to distinguish ordinary use from the small number of decisions capable of imposing extraordinary risks on people who never consented to them.

Should it be built at all?

Licence the most dangerous capabilities, not every AI product. Above clearly defined thresholds, developers should need permission before deployment and should declare exceptionally large training runs. Thresholds must reflect what a system can do, not merely its size. Requirements should be proportionate, publicly defined and regularly reviewed, so that regulation does not become a barrier designed by large incumbents against smaller competitors and universities.

Put the burden of showing safety on the developer. Before release, require a documented safety case: capabilities, intended uses, foreseeable misuse, failure modes, available safeguards, evidence from tests and conditions requiring suspension. Independent laboratories should test for deception, attempts to gain unauthorised resources or permissions, avoidance of oversight, persistence after shutdown instructions and assistance with catastrophic misuse. Regulators should receive full findings; the public should see as much as can safely be disclosed. Testing methods must evolve as capabilities change. An assessment passed before a model acquires new tools or greater autonomy cannot be treated as a permanent certificate of safety.

Assess the deployment as well as the model. Drafting a letter is not equivalent to managing a power grid. Risk depends on what systems the model can reach, what authority it has, how long it can operate unattended and whether its actions can be reversed.

What is it allowed to reach?

Keep meaningful human control over critical systems. Water supply, electricity, transport signalling, medical equipment and financial clearing require accountable human decisions for consequential actions. That human must understand the proposed action, have time to challenge it and be able to refuse; clicking 'approve' without those conditions is not meaningful supervision. Irreversible decisions may need two independent approvals. AI should not autonomously initiate nuclear weapons use or select and attack people without meaningful human control.

Separate advice from execution. Where failure could kill people at scale, control networks should be isolated from general-purpose AI. Elsewhere, grant agents only the permissions necessary for a specific task, for a limited time, with limits on autonomous actions. They must not be allowed to acquire new credentials, replicate themselves, move into other environments or disable their own supervision without explicit authorisation.

Keep essential services capable of operating without AI. Manual procedures, trained staff, offline backups and tested recovery plans must survive. A shutdown mechanism is of little use if society has become unable to run the service after shutting it down.

How would anyone know?

Certify the places where models are tested. Evaluation environments should meet enforceable isolation and security standards. If safety protections are disabled for a test, that exception should be recorded and controlled. Containment must be demonstrated, with independent kill mechanisms, tamper-resistant logs and monitoring that alerts when something goes wrong. Outside access matters: independent researchers identified a monitoring weakness during a 2026 pilot that OpenAI subsequently patched.

Watch after release, not just before it. High-risk deployments should undergo continuing assessment, with automatic suspension at defined thresholds. Serious incidents and near misses must be reported promptly to an authority and entered in an appropriately public register, using compatible international definitions. Affected organisations and people must receive direct notice, rather than having to discover a breach through the news. A public register should explain what happened and what was done about it, while withholding details that would hand attackers a usable exploit. Near misses deserve attention precisely because they offer lessons before lives are lost.

Protect the people who raise the alarm. Whistleblowers, contractors and independent evaluators need legal protection when they disclose concealed safety failures. Non-disclosure agreements must not silence them. Significant model theft should be reported, and developers and major computing providers should meet security and record-keeping standards appropriate to the risks they enable.

Give regulators the means to act. Independent funding, technically competent inspectors and access to evidence are essential. A levy set in law can help finance oversight without letting the industry control it. Publicly funded research must also investigate how systems can be understood, controlled and safely stopped. Inspectors should be trained to examine models, evaluation arrangements and the networks around them, and moved between assignments to reduce the risk of becoming too close to the organisations they oversee.

When it goes wrong

A serious incident should trigger an independent investigation, not an inquiry controlled by the developer. Investigators need legal powers to preserve logs, inspect documentation, question those involved and publish findings where that does not create another danger.

Countries also need emergency plans agreed before an incident: who leads, who warns infrastructure operators, who may suspend a system and how international partners are notified. Emergency powers must be tightly defined, time-limited and subject to courts and parliament. Fear of AI cannot become an excuse for unchecked government power. Victims also need a clear route to information, remedies and compensation while investigators establish the facts.

Who answers for it?

Responsibility should follow control and conduct. Depending on the circumstances, a developer, deployer, infrastructure provider or operator may have caused or failed to prevent harm. Law should distinguish negligence, recklessness and deliberate misuse. An accident despite reasonable precautions is not the same as knowingly disabling safeguards to beat a competitor to market. Criminal accountability should attach to culpable conduct, not merely to the fact that an AI system was involved. Outsourcing evaluation or containment must not let a developer wash its hands of responsibility.

There is already a useful parallel in the UK's Automated Vehicles Act 2024, which establishes responsibilities for authorised self-driving entities and operators rather than assuming the person in the passenger seat controls the vehicle. AI legislation likewise needs to identify who answers when autonomous systems cause harm.

For the most capable systems, require a named senior officer responsible for containment and adequate insurance or a financial bond. Regulators need meaningful fines, suspension powers and the ability to require withdrawal. Prompt reporting and cooperation should reduce penalties where appropriate; concealment, misleading investigators and retaliation against safety staff should make them worse. Otherwise, the law will reward silence.

Countries working together

Neither software nor a containment failure respects borders. National regulators should exchange warnings about serious incidents, stolen models and emerging dangerous capabilities, with mandatory notification when a system affects another country.

We have precedents for international oversight. The International Atomic Energy Agency inspects aspects of nuclear activity; the Organisation for the Prohibition of Chemical Weapons oversees a treaty banning chemical weapons. AI will be harder to verify because software can be copied and moved almost instantly. That makes specific, checkable duties more valuable than sweeping promises.

The UN established an Independent International Scientific Panel on AI and a Global Dialogue in 2025, but the panel has no inspection or enforcement power. The EU AI Act already establishes some incident-reporting duties. These are starting points, not a complete international framework. Nor must governments wait for a perfect world treaty before acting: they can begin with shared reporting definitions, confidential warnings between regulators and agreements to recognise credible independent tests.

Governments should negotiate common testing standards, cross-border reporting, access for inspectors and narrow prohibitions on especially dangerous autonomous uses. A shared minimum standard would also stop companies and countries being penalised commercially for taking safety seriously. This is not a call to ban AI; it is a call to remove the advantage gained by ignoring its risks.

The practical difference would be straightforward: an organisation releasing or deploying a powerful system would know its limits; an inspector could verify the safeguards; victims would be informed; investigators could establish responsibility; and governments could learn from failures before they spread. The technology might be identical. The incentives surrounding it would not.

The "Do Something about it" Part

None of these decisions requires you to become an AI engineer. They require legislators willing to ask difficult questions, pass enforceable laws and fund independent oversight. Procurement decisions also matter: governments can insist on tested systems when buying technology for public services.

There are three things you can do before the next election:

Find out what is being proposed. Look for policies on pre-release testing, accountability, incident reporting, whistleblower protection, emergency planning and limits on autonomous access to critical infrastructure. Ask whether essential services where you live can still operate if their AI systems are switched off. Bring the questions to unions, professional bodies, community groups and public meetings.
Find out where candidates stand. My article Where Do They Actually Stand? provides prompts for investigating a candidate with an AI chatbot. Enable web search, seek voting records and published statements, and check the links yourself. No evidence of a position is a reason to ask the candidate directly, not to invent one for them.
Ask one question and publish the answer. For example: “Should an AI developer face legal consequences when its model breaks containment and harms another organisation?” Or: “Should serious incidents and near misses be reported publicly?” Ask in writing, at a meeting or on social media. Share the answer, or the fact that none was received. Then check whether the candidate follows through. A promise to make your country an 'AI leader' is not an answer to a question about safety, accountability or public rights. Ask for a commitment specific enough that you will be able to check it later.

Then vote. Choose representatives whose record and proposals best match the protections you want, and keep holding them to those commitments afterwards. Voting gives you a say in who holds power; organising and following up give that vote weight between elections.

We cannot vote risk out of existence. We can decide who may impose it on others, what precautions they must take, who investigates failures and who pays when things go wrong. Those decisions should not be left to the handful of people building the technology.

Send this article to someone who thinks AI is harmless, someone who thinks disaster is inevitable, or a candidate who has never been asked about AI safety. Ask which proposals they would support.

Arguing with a technology changes little. Changing the rules governing the people who control it — and voting for representatives prepared to do that — can change a great deal.

Follow This, That and (Maybe), the Other:

Search This Blog

TT(O)M