---
title: "How to talk about AI regulation and independent oversight"
url: "https://commonsignals.org/how-to-talk-about-ai-regulation"
published: 2026-11-17
updated: 2026-11-17
publisher: "Common Signals"
---

# How to talk about AI regulation and independent oversight

17 November 2026 · Ben Matthews

Three dos and a don't for making the case for rules with teeth, and the words to avoid while you do it.

This guide is written for communicators and campaigners drafting messages now. Each item gives a case, one verifiable proof point, and a ready-to-use script; methodology sits in the Sources.

This is the second of seven issue guides expanding the short versions in [How to communicate about AI safety](https://commonsignals.org/how-to-communicate-about-ai-safety#2-regulation-and-independent-oversight). It covers the case for an independent regulator, testing before release, and binding rules for the most powerful AI models.

## Does the public want AI regulated?

Yes, so you do not need to convince people that AI needs rules. In [Diffusion's survey](https://diffusion-beta.netlify.app/assets/Diffusion-UK-AI-Compass-Report.pdf) of 2,911 UK adults, 85% said the UK needs stronger laws to make AI safe, and 10% would rather trust companies to regulate themselves.

The [Ada Lovelace Institute](https://www.adalovelaceinstitute.org/policy-briefing/great-expectations/) found that 89% of UK adults think it is important that AI is regulated independently, and 89% that AI products shouldn't be rolled out until they are proven safe, even if that slows things down.

In the US, [79% of adults](https://www.americanson.ai/) said in October 2026 that dealing with AI harms should be a priority for government.

The difficulty is trust. 80% of UK adults distrust the government's ability to control AI (Diffusion), and 84% are concerned that the government will put the needs of the technology sector ahead of the public when it regulates AI (Ada Lovelace Institute).

People want rules and doubt the people who would write them. That distrust may be why the word "independent" is likely to carry weight, but this is an inference. The survey does not test whether people read "independent" as independent of government.

The evidence points to three things to bring to every conversation: an independent body with named powers, a standard people already accept for other products, and the gap between what was promised and what the law requires.

## 1. Say "independent regulator", and spell out its powers

Independent regulators are viewed more favourably than any other institution in Diffusion's survey, and in its dial test, scores for the regulator message rose while it described what the body could do. Name the body and its powers in the same breath.

**The case study:** In Diffusion's dial test, UK adults rated a message calling for an independent regulator with powers to inspect, audit and shut down dangerous AI systems. In every one of Diffusion's six mindsets, scores rose by 11 to 18 points during the passage where the message described those powers (the 30th to 42nd seconds). A dial test cannot separate the powers from the message's emotional build or close. Over the same 12 seconds, the gap between the most and least supportive mindsets narrowed from 8.1 points to 2.7. The advice to spell out the powers rests on the message's overall strength and the appetite evidence below, not on proof that the powers caused the gain.

Separately, and not from a dial test, news audiences reach for the same idea unprompted. Under the [Amanpour interview](https://commonsignals.org/problem-they-created-rogue-ai-agents-openai-huggingface-amanpour-khlaaf) in which Heidy Khlaaf framed the incident involving a swarm of OpenAI research agents as a question of accountability, 12.4% of comments called for watchdogs, oversight, prosecution or action by Congress, more than double the roughly 5% under earlier videos in our series (Observed). This shows appetite for the idea, not that the powers drove the dial-test gain.

**The evidence:** Independent regulators scored net +26 in Diffusion's survey (net score is the share favourable minus the share unfavourable), against -24 for AI technology companies. The regulator message finished at [72.7 out of 100](https://diffusion-beta.netlify.app/articles/uk-compass-report), second of the five messages tested. By contrast, the AI Security Institute, asked about by name, scored net -8, well below independent regulators in general. The powers themselves have wide support: 89% of UK adults say government and regulators should have the authority to halt the use or sale of any AI system found to cause harm ([Ada Lovelace Institute](https://www.adalovelaceinstitute.org/policy-briefing/great-expectations/), Tested).

**Put it into practice:** "An independent watchdog that can test AI before it's released, investigate when things go wrong, and switch off dangerous systems." Set that against "better AI governance", which gives people nothing to picture.

**Rule of thumb:** Lead with what the body does, and treat its name as secondary. If you name the AI Security Institute, put its function first and the name after. The name alone scores net -8, and the survey can't tell whether that reflects low awareness or distrust. Pairing the name with its function is a hedge: it may help if people simply don't know the body, but it won't fix distrust.

## 2. Anchor the case in "proven safe before release"

Nine in ten UK adults already back the goal: AI proven safe before it is released. The same survey found that 82% support mandatory safety testing, though the briefing reports no question on whether a regulator should approve models before release. The case for that mechanism is the argument that follows. It borrows a standard people accept for medicines, aircraft and lifts, and it puts the burden of proof on the company: the company shows its product is safe, and the regulator checks.

**The case study:** The UK's [AI Security Institute](https://www.aisi.gov.uk) already [tests some powerful models before release](https://www.eweek.com/news/ai-predeployment-testing-emea-uk/), including American companies' models, but only when the companies agree. It has no power to require a test or to stop a release. In September 2026, Parliament's Joint Committee on Human Rights [recommended](https://publications.parliament.uk/pa/jt5902/jtselect/jtrights/160/report.html) changing that. Developers of powerful models would have to submit them for testing, the Institute would publish its findings before release, and an independent regulator could block a release or order a system off the market.

**The evidence:** 89% of UK adults agree that AI products and services should not be rolled out until they are proven safe, even if this slows things down ([Ada Lovelace Institute](https://www.adalovelaceinstitute.org/policy-briefing/great-expectations/), 1,928 adults, September 2025).

**Put it into practice:** The UK government's reply to the committee was due by 14 November 2026, so the decision is live. Check whether the reply has been published before you use this line. "Nine in ten of us want AI proven safe before it's released. Right now, the companies decide whether an independent tester sees it first."

**Rule of thumb:** Borrow a standard people already trust. Medicines, aircraft and lifts are checked before anyone uses them, and nobody expects the manufacturer to be the only judge.

## 3. Name the gap between promises and law

Put what was promised next to what happened, date both, and let the reader draw the conclusion.

**The case study:** At the [AI Seoul Summit](https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024) in May 2024, 16 AI companies signed voluntary safety commitments. ControlAI's [Artificial Guarantees](https://controlai.org/artificial-guarantees) tracks promises like these against what the companies did next. OpenAI committed to publish reports for all significant new model releases. OpenAI then launched GPT-4.1 with no plans for a system card. Google DeepMind made the same commitment. Google DeepMind then released what it called its "most intelligent AI model" without a model card, and PauseAI [protested](https://pauseai.substack.com/p/we-held-the-largest-ai-safety-protest) outside Google DeepMind.

The government made its own promise. In the July 2024 King's Speech, the government said it would seek "appropriate legislation" for developers of the most powerful models. No government bill has followed, and in October 2026 MLex reported that the government had still not decided on [AI laws and the Institute's powers](https://www.mlex.com/mlex/articles/2537666).

![ControlAI tracker entry setting an OpenAI commitment against the GPT-4.1 launch](https://commonsignals.org/assets/toolkit/S05-controlai-artificial-guarantees-gpt41.png)

What was promised, and what happened. ControlAI. Captured 17 November 2026.

**The evidence:** In [Diffusion's dial tests](https://diffusion-beta.netlify.app/articles/uk-compass-report), the highest-scoring messages named a specific actor, a tangible consequence and a clear remedy. Messages that offered hope without anyone accountable did well in the moment and faded on reflection. No tested message used the promise-versus-outcome structure, so the gap framing is an application of the accountable-actor finding, not something the dial tests measured directly.

**Put it into practice:** "OpenAI promised a safety report for every significant release. Then it launched GPT-4.1 without one." Or: "In 2024 the government promised legislation for the most powerful AI models. There is still no government bill."

**Rule of thumb:** Use the company's or the government's own words for the promise, with a date and a link. A promise in their words is harder to dispute than your summary of it.

## And a don't: industry's vocabulary

Don't make the case in the language of "light-touch", "pro-innovation" or "responsible self-governance". Those phrases point towards the voluntary approach that 10% of UK adults back (Diffusion, Tested), and they carry a policy history: the UK government's 2023 framework was called [A pro-innovation approach to AI regulation](https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach).

The innovation framing also fades on reflection. In Diffusion's dial tests, a message about innovation and growth had the highest early score of the five, then lost ground: it was the only message that several mindsets rated lower on reflection. It finished fourth of five, at 67.1, a modest gap of 5.6 points behind the independent-regulator message at 72.7. It named no one responsible (Tested). If you adopt the language, you inherit the argument.

## Example messages

**Full version:**

"We expect medicines, aircraft and lifts to be checked by independent inspectors before anyone uses them. Nine in ten of us want AI held to the same standard: proven safe before it is released. Right now, the companies building the most powerful AI decide for themselves whether it is safe enough, and the UK's AI Security Institute can test their models only if they agree. An independent public watchdog with the legal power to test these systems before release, investigate failures and switch off dangerous ones is what people already expect. It is time the law caught up."

**Short version (for social):**

"Nine in ten of us want AI proven safe before it's released. Right now, the companies decide whether an independent tester sees it first. Britain needs an independent watchdog with the power to test AI and switch off dangerous systems."

## Objections to prepare for

Write your answers before you need them. Two of these are the government's own stated position, so expect to hear them from officials as well as the public.

### "Regulation will kill innovation."

**The trap:** Arguing about growth on the other side's terms.

**The pivot:** Point to standards people already rely on. "We test medicines before they're sold, and nobody thinks that killed medicine. Safety checks are what let people trust a product enough to use it."

**Optional second pivot, for audiences worried about foreign dependence:** Shift to who the UK relies on. Framing AI as dependence on a few US tech companies raised the share rating AI sovereignty very or extremely important from 53% to 63% (Diffusion, Tested). That shows the frame's appeal, not how people answer the innovation objection. See the [issue guide on AI, war and national security](https://commonsignals.org/how-to-communicate-about-ai-safety#6-ai-war-and-national-security).

### "Why would I trust the government with this?"

**The trap:** Defending the government. 80% of UK adults distrust its ability to control AI (Diffusion).

**The pivot:** Agree, and make the case for a regulator independent of ministers and of the companies, which publishes what it finds. "You shouldn't have to trust ministers or the companies. That's why the checks need an independent body that publishes its results."

### "We already have regulators for this."

**The trap:** A debate about which existing regulator covers what.

**The pivot:** This is the government's current position: regulate AI where it is used, through existing regulators. The Joint Committee on Human Rights found that this makes the organisations using AI "the primary holders of responsibility" and leaves developers largely outside the rules. "Existing regulators deal with AI once it's in use. None of them can test a powerful model before it's released, or stop the release."

### "You can't just switch AI off, and it's built abroad anyway."

**The trap:** Promising a kill switch as the whole answer.

**The pivot:** This is the government's position. In September 2026 it rejected Lords amendments that would have given ministers last-resort powers to shut down AI systems, and the Cabinet Office argued that the UK can't [simply turn AI off](https://www.legal500.com/intelligence/united-kingdom/technology/will-the-uk-change-its-course-on-ai-regulation), and that blocking access here wouldn't stop systems being built or misused elsewhere. Move the ask to testing before release, which applies to anything released here, whoever builds it. "A medicine made in America still needs approval before it's sold here. AI can meet the same rule. The AI Security Institute already tests American companies' models; it just can't insist."

## Questions people ask

### What do I say when people think no government will act?

Give them a decision with a name and a date. Under the Amanpour interview, 4.7% of comments named the US administration as the reason nothing would happen (Observed). Cynicism like this is a reason to make the ask smaller and more specific. A named decision gives people somewhere to put the concern, which is what [research on fear and efficacy](https://commonsignals.org/how-to-communicate-about-ai-safety#2-pair-every-threat-with-something-people-can-do) recommends. In the UK, that decision is the government's reply to the Joint Committee on Human Rights.

### How strong is the evidence behind these rules?

It varies by rule, and the Sources label each figure. Tested means a single experiment or representative survey: the 85%, 89% and 79% opinion figures and Diffusion's dial-test scores are Tested. Observed means a pattern in a self-selected comment thread: the 12.4% of Amanpour comments calling for oversight is Observed. Build a message on the Tested findings and treat the Observed ones as [hypotheses to try](https://commonsignals.org/how-to-communicate-about-ai-safety#how-we-grade-ai-safety-evidence).

### Does this advice apply outside the UK?

Partly. Most of the support figures are British, though 79% of American adults also want AI harms treated as a government priority, and Diffusion ran its dial test with UK adults only. Rules 2 and 3 name UK institutions and a UK decision. Elsewhere, keep the structure: an independent body with named powers, a standard people already accept, and a dated promise set against what happened. Then name your own regulator and a decision that is live now.

## Sources

- Diffusion, [The UK AI Compass](https://diffusion-beta.netlify.app/assets/Diffusion-UK-AI-Compass-Report.pdf), survey of 2,911 UK adults, December 2025, and [dial-test results](https://diffusion-beta.netlify.app/articles/uk-compass-report).
- Ada Lovelace Institute, [Great (public) expectations](https://www.adalovelaceinstitute.org/policy-briefing/great-expectations/), 1,928 UK adults, fieldwork September 2025, published December 2025.
- Americans on AI, [biweekly US tracker](https://www.americanson.ai/), Athena Insights and NORC, surveys fielded 8 to 26 October 2026.
- Joint Committee on Human Rights, [Human Rights and the Regulation of AI](https://publications.parliament.uk/pa/jt5902/jtselect/jtrights/160/report.html), HC 160 / HL Paper 56, 14 September 2026.
- MLex, [UK AI laws and security institute powers remain undecided](https://www.mlex.com/mlex/articles/2537666), 14 October 2026.
- Legal 500, [Will the UK change its course on AI regulation?](https://www.legal500.com/intelligence/united-kingdom/technology/will-the-uk-change-its-course-on-ai-regulation), 29 September 2026.
- eWEEK, [UK weighs statutory AI pre-deployment testing](https://www.eweek.com/news/ai-predeployment-testing-emea-uk/), 6 August 2026.
- UK government, [Frontier AI safety commitments, AI Seoul Summit 2024](https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024).
- UK government, [A pro-innovation approach to AI regulation](https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach), 2023.
- ControlAI, [Artificial Guarantees](https://controlai.org/artificial-guarantees).
- PauseAI, [report on its Google DeepMind protest](https://pauseai.substack.com/p/we-held-the-largest-ai-safety-protest), 2025.
- Message test details (Tested): the independent-regulator message scored 72.7/100 in [Diffusion's dial tests](https://diffusion-beta.netlify.app/articles/uk-compass-report), second of five, and gained 11 to 18 points in every mindset during the passage describing its powers. The innovation and growth message finished fourth, at 67.1.
- Sovereignty split test (Tested): the share rating AI sovereignty extremely or very important was 53% with neutral wording, 63% with a US-dependence frame and 55% with a China frame ([Diffusion](https://diffusion-beta.netlify.app/articles/uk-compass-report)).
- [Comments on the Amanpour interview with Heidy Khlaaf](https://commonsignals.org/problem-they-created-rogue-ai-agents-openai-huggingface-amanpour-khlaaf) (Observed): 12.4% of comments used regulation-type language, against roughly 5% under earlier videos in the series, and 4.7% named the US administration as the reason nothing would happen.

## Related research

- [OpenAI and shut? What an independent report did to the marketing-stunt argument](https://commonsignals.org/openai-and-shut-metr-hugging-face)
- ["A problem they created": how a news audience received the accountability version of the OpenAI and Hugging Face incident](https://commonsignals.org/problem-they-created-rogue-ai-agents-openai-huggingface-amanpour-khlaaf)
- ["Neither company is acting responsibly": how X received Jacob Coxon's resignation from Anthropic](https://commonsignals.org/coxon-resignation-x-analysis)
- ["Can't we just unplug it?": The Diary of a CEO AI debate analysed through 26,000 comments](https://commonsignals.org/cant-we-just-unplug-it-doac-ai-debate-comment-analysis)
- [Findings so far: what four comment-thread studies say about explaining AI risk](https://commonsignals.org/findings)
- [How to communicate about AI safety](https://commonsignals.org/how-to-communicate-about-ai-safety)
- [AI safety comms reading list](https://commonsignals.org/ai-safety-comms-reading-list)
- [Methodology: how we code comment threads, and what our studies can and can't tell you](https://commonsignals.org/methodology)
