iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
“Doomer” and “booster” are informal labels, not memberships in two organisations. In the current AI argument, doomers are generally associated with warnings about severe or even catastrophic outcomes from advanced systems. Boosters, also called accelerationists in some online discussions, generally prioritise rapid development and are more sceptical that existential-risk warnings justify slowing progress. Many influential people fit neither stereotype: they support major AI benefits while arguing that specific risks require testing, oversight and limits.
The useful question is not which team a person belongs to. It is what risk they are discussing, what evidence they accept, how much uncertainty they tolerate and which safeguards they would require before deployment.
What “doomer” and “booster” mean in the AI debate
The terms became popular through online debate and are used as shorthand in news coverage rather than as formal political or professional affiliations. A person labelled a doomer may be warning about present misuse, future loss of control, or both. A person labelled a booster may favour fast capability growth while still supporting some regulation. Neither label reliably predicts a person’s views on every AI policy.
The central disagreement is often narrower than the labels suggest. Most participants can see substantial benefits from better systems. The harder question is whether those benefits justify accepting a high level of technological risk, and how much evidence governments and companies should demand before releasing increasingly capable models.
#1 Best Overall
“Existential risk” usually means a possibility of permanent, civilisation-scale damage, such as humans losing control over systems that can pursue goals more effectively than institutions can respond. The same debate also covers less speculative harms: fraud, cyberattacks, discriminatory decisions, unsafe medical advice, labour-market disruption and synthetic media. Treating all of these as one probability obscures important differences.
Who are the AI risk-warning voices?
A September 15, 2026 Euronews explainer places several prominent researchers and technology leaders on the risk-warning side of the discussion. That is a description of how the article presents their public positions, not proof that they share a single programme.
| Figure | How the Euronews account characterises the position | Important qualification |
|---|---|---|
| Dario Amodei | Presented as a prominent voice warning about serious risks from advanced AI. | The article notes accusations of hypocrisy because Anthropic, the company he leads, also develops frontier systems. A warning about risk does not mean opposition to all development. |
| Geoffrey Hinton | Named among high-profile researchers who have raised concerns about advanced AI. | The label describes a public emphasis on risk, not a complete statement of his views on every technical or regulatory question. |
| Yoshua Bengio | Included in the group calling attention to potentially severe consequences. | Being grouped with other warning voices does not establish identical timelines or preferred interventions. |
| Max Tegmark | Associated with the 2023 letter calling for a six-month pause in development of the most advanced systems. | The letter was a specific proposal at a specific point in capability development, not evidence that every risk-aware researcher supports a permanent moratorium. |
The common thread is a demand to take tail risks seriously even when their probability is disputed. This can lead to calls for evaluations before release, limits on dangerous capabilities, more transparency and international coordination. It does not automatically imply that useful AI research should stop.
Who are the AI boosters or accelerationists?
The same Euronews article uses “booster,” and sometimes “accelerationist,” for voices that favour rapid development or reject broad existential-risk claims. The people in this group also differ substantially from one another.
| Figure | How the Euronews account characterises the position | What not to infer |
|---|---|---|
| Marc Andreessen | Described as opposing regulation that would slow AI development. | This does not establish opposition to every safety practice or a shared position with other people called boosters. |
| David Sacks | Also described as opposed to regulation in the article’s account. | A stance against particular rules is not the same as a technical claim that AI has no risks. |
| Jensen Huang | Reported as dismissing extinction claims while supporting a unified federal framework and industry peer review. | His position demonstrates why the two-team framing is inadequate: scepticism about extinction can coexist with support for formal oversight. |
| Yann LeCun | Reported as calling existential-risk warnings overblown. | The description concerns the scale or plausibility of existential scenarios, not a declaration that current AI systems are harmless. |
Accelerationist arguments often focus on the cost of delay: slower scientific discovery, weaker national competitiveness, missed productivity gains and less access to tools that could improve health, education or climate work. Critics respond that a race can encourage companies and states to release systems before testing is adequate, spreading failures faster than institutions can correct them.
The cautious middle: innovation with a risk ceiling
Sam Altman’s September 2026 remarks illustrate a position that does not fit neatly into either camp. He said the industry should not accept excessive technological risk merely because AI’s benefits are large, and he rejected both blind optimism and doomerism. He also argued that competition by itself is not a reason for rash decisions.
In remarks at the United Nations Security Council in September 2026, Altman said: “The industry must not accept too much technological risk just because the benefits are too great and that they feel too important to slow down.” This is a governance argument, not a claim that every risk forecast is correct. It accepts the value of continued progress while setting a boundary around risks that should not be traded away for speed.
A middle position can therefore support frontier research, deployment of lower-risk systems and strong commercial incentives while requiring evidence and controls for systems with dangerous capabilities. It may favour staged releases, independent testing and incident reporting rather than either unrestricted acceleration or a blanket pause.
Why experts disagree about existential risk
Disagreement is not explained only by personality or ideology. The International AI Safety Report 2026 identifies different kinds of evidence and uncertainty that lead reasonable analysts to different conclusions.
They are often talking about different time horizons
Some analysts concentrate on harms already observed: synthetic media used for deception, cybersecurity vulnerabilities, privacy failures or unsafe outputs in high-stakes settings. Others focus on systems that do not yet exist and ask what could happen if models gain much stronger planning, autonomy, persuasion or ability to operate software and machines. Evidence for the first category can come from incidents and measured tests. Evidence for the second often depends on models of future capability, controlled laboratory studies and theoretical analysis.
They assign different weight to uncertain capability growth
The UK government’s 2024 interim international report says researchers disagree about how quickly capabilities will advance. A person who expects rapid progress may favour safeguards before a capability appears in public systems. Someone who expects slower or uneven progress may see early restrictions as premature and potentially harmful. Both are making forecasts, not reporting settled facts.
They have different expectations about safeguards
The same interim report says views vary over how effective technical and policy measures will be. If monitoring, access controls, model evaluations and secure deployment are expected to work well, a society may tolerate more experimentation. If those measures are expected to fail under competitive pressure, stronger limits may be justified earlier.
They disagree about what society will do
Risk estimates include assumptions about government response, company behaviour, public adoption and international cooperation. A scenario that seems unlikely under coordinated oversight can look more plausible if states race independently and companies conceal failures. Conversely, a warning may appear overstated to someone who expects institutions to adapt quickly.
They choose different precaution thresholds
The 2026 report describes an “evidence dilemma.” Acting before evidence is conclusive can produce ineffective or harmful interventions, lock in poor standards or deny benefits. Waiting for conclusive evidence can leave people exposed while a dangerous capability spreads. The disagreement is therefore partly empirical and partly ethical: how much possible harm should justify preventive action?
They face different national circumstances
Risks do not arrive in the same way everywhere. Adoption rates, digital infrastructure, labour markets, legal systems and state capacity vary by country. A policy that is workable for a country with strong testing institutions may be difficult for one without them. The International AI Safety Report 2026 warns that the effects of AI risks can differ across national and institutional contexts.
What the strongest current evidence actually shows
The International AI Safety Report 2026 was prepared by more than 100 independent experts. Its advisory panel included nominees from more than 30 countries and intergovernmental organisations. The report reviews evidence available through December 2025, making it a substantial 2026 synthesis rather than a live count of every subsequent incident.
Current harms have firmer empirical support
The report says there is robust empirical evidence for harms including AI-generated media and cybersecurity vulnerabilities. These claims can be studied through documented incidents, security testing and observed system behaviour. Their existence does not depend on assuming a future superintelligence.
Frontier scenarios carry more uncertainty
Some risks associated with future capabilities rely on modelling, controlled laboratory studies and theoretical analysis. Such methods can reveal important failure modes, but they cannot establish a precise probability that a particular catastrophe will occur. They are evidence for preparation and investigation, not a settled forecast.
There is no authoritative “P(doom)” number
“P(doom)” is online shorthand for a person’s estimated probability of an existential catastrophe. The phrase appears in public discussions and in the title of a 2025 arXiv paper about why experts disagree, but the available material does not establish a consensus percentage. Treating a quoted personal estimate as a scientific measurement confuses an individual judgement with a validated statistic.
Nothing about the future is predetermined
The UK government-published interim international report states: “But nothing about the future of AI (artificial intelligence) is inevitable.” The point is practical. Choices about system design, deployment, regulation, investment and international coordination can change which benefits and harms occur.
Compare positions by the questions they answer
Instead of sorting people into two teams, compare their answers on five policy dimensions.
| Question | Risk-warning emphasis | Booster or accelerationist emphasis | Middle or governance-focused approach |
|---|---|---|---|
| What risk matters most? | Low-probability, high-consequence loss of control and severe misuse, alongside current harms. | Immediate benefits and existing problems caused by slow adoption or excessive restrictions; existential claims may be viewed as overstated. | Separate present, measurable harms from uncertain frontier scenarios and address each with proportionate controls. |
| What counts as evidence? | Models, red-team results, controlled experiments and theoretical failure analysis can justify preventive action. | Observed incidents and reproducible real-world failures deserve more weight than speculative scenarios. | Use both empirical and forward-looking evidence, with clear disclosure of uncertainty. |
| When should rules begin? | Before dangerous capabilities are widely deployed, because prevention may be easier than recovery. | After evidence of concrete harm, to avoid blocking beneficial innovation. | Use staged thresholds, testing and reversible deployment where possible. |
| Who should govern? | Independent evaluators and binding public oversight are needed to counter competitive pressure. | Voluntary commitments, technical progress and limited regulation may be more adaptive. | Combine company safety work with independent access, common standards and enforceable rules for high-risk systems. |
| How should benefits and burdens be shared? | Prevent safety, labour and security costs from falling on people with no control over deployment. | Expand access quickly so that productivity and scientific gains are not concentrated among incumbents. | Pair safety measures with international access and avoid widening gaps between countries. |
Policy proposals between a pause and a free-for-all
A September 22, 2026 joint diplomatic statement hosted by the Netherlands illustrates the practical ground between the poles. It calls for transparent safety protocols, mandatory pre-deployment testing, independent evaluation with sufficient access, coordinated standards and shared reporting of serious safety incidents. It also says risk measures should not widen the gap between countries’ access to AI benefits.
These are proposals in a diplomatic statement, not a global rulebook already in force. Their value is that they identify concrete mechanisms that can be debated: test a system before release, let an evaluator inspect enough of it to make a meaningful judgement, use compatible standards across borders and tell other authorities when severe failures occur.
OpenAI’s 2025 recommendations similarly call on frontier laboratories to share safety research and consider shared safety principles and control-evaluation standards. Those recommendations are the company’s published position and forecast, not independent scientific consensus. They show how a company can argue for coordination while continuing to develop advanced systems.
Best Value
What pre-deployment testing can and cannot do
Testing can reveal dangerous capabilities, unreliable refusals, cyber vulnerabilities and unexpected interactions before a model reaches millions of users. It cannot prove that every future use will be safe, especially when users combine a model with tools, data or autonomous action that were not present in the test environment. Test results therefore need scope, dates, access conditions and limitations attached to them.
Why independent evaluation matters
A company’s internal evaluation may be technically sophisticated but still constrained by commercial incentives, limited access to training details or a narrow view of downstream use. Independent evaluation can challenge assumptions, reproduce findings and compare systems under common methods. Independence only helps if evaluators receive enough access to inspect the relevant model, tools and documentation.
Why incident reporting is a shared-interest measure
Serious incidents are costly sources of information. Sharing them can help other developers patch similar weaknesses before they are exploited. Reporting rules need clear thresholds and protections for sensitive security details; otherwise companies may either over-report trivial events or conceal important ones.
How to read a claim about “doomers” or “boosters”
- Identify the exact risk. Is the speaker discussing fraud happening now, job displacement over the next decade, or a hypothetical loss of control from a future system?
- Check the evidence type. Ask whether the claim relies on an observed incident, a benchmark, a controlled experiment, a model or a theoretical argument.
- Separate probability from severity. A low-probability event can deserve preparation if the damage would be extreme, but severity alone does not establish likelihood.
- Look for assumptions about safeguards. Find out whether the argument assumes effective monitoring, international cooperation, secure infrastructure or rapid institutional adaptation.
- Ask what intervention is actually proposed. “Take risks seriously” could mean a voluntary audit, a licensing rule, a temporary pause or a permanent ban. Those are not interchangeable.
- Check who bears the cost. A rule can reduce one danger while increasing concentration, denying poorer countries access or shifting work and infrastructure burdens onto less powerful groups.
- Distinguish a person’s words from a label. The terms are usually applied by commentators. Read the underlying proposal before deciding where someone stands.
What the debate means for people who use AI
For ordinary users, the disagreement is not a reason to wait for a final verdict on AI. It is a reason to match trust and oversight to the consequences of a task. Treat generated text as a draft when accuracy matters, verify claims in medical, legal or financial contexts, protect confidential data and keep a human decision-maker responsible for high-impact choices.
For organisations, the practical response is to document which systems are used, what data they receive, who can approve outputs and how incidents are escalated. Higher-risk uses warrant stronger access controls, testing against realistic failure modes and a way to suspend or roll back a deployment. These steps remain useful whether a company’s long-term forecast is optimistic or pessimistic.
For policymakers, the challenge is to regulate demonstrated harms without pretending that uncertain forecasts are settled facts. Rules can be updated as evidence improves, provided they include review dates, transparency requirements and ways to correct ineffective interventions. International coordination matters because model development, data flows and misuse cross borders even when laws do not.
The bottom line on doomers versus boosters
The global AI risk debate is not a contest between people who hate technology and people who love it. It is a disagreement about evidence, timing, acceptable uncertainty and who should control powerful systems. Risk-warning voices emphasise consequences that may be difficult to reverse. Booster voices emphasise the costs of delay and doubt that existential scenarios justify sweeping restrictions. The cautious middle accepts that both benefits and dangers are real and argues for progress bounded by testing, independent scrutiny, shared standards and clear limits on unacceptable risk.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
