What are AI researchers worried about?

Based on a survey of over 4,000 AI researchers, this paper reveals that contrary to public and media narratives dominated by existential threats, researchers prioritize immediate sociotechnical risks and show significant convergence with public opinion on risk assessment, suggesting a need for collaborative dialogue focused on mitigating present-day harms rather than speculating on future catastrophes.

The Gist

Imagine the world of Artificial Intelligence (AI) as a massive, bustling construction site. Right now, the public debate is like a noisy crowd gathered behind a fence, listening to a few loud megaphones held by a handful of wealthy developers and famous scientists. These megaphones are shouting two extreme things: either "This machine will save us all!" or "This machine will destroy humanity tomorrow!"

But the people actually doing the work—the thousands of engineers and scientists laying the bricks and mixing the cement—have a very different, much quieter conversation going on inside the site.

This paper is like a team of sociologists sneaking onto that construction site to ask the workers, "Hey, what are you actually worried about?" They surveyed over 4,000 AI researchers to see if their fears match the scary headlines or the utopian dreams we see in the news.

Here is what they found, translated into everyday language:

1. The "Sci-Fi" vs. "Real Life" Gap

The Headline: The media loves to talk about "Existential Risk"—the idea that AI will become a super-intelligent robot that wakes up, decides humans are useless, and ends the world (think The Terminator or SkyNet).

The Reality: When asked, "What is your biggest worry?", only 3% of AI researchers said they were scared of the robot apocalypse.

• The Analogy: It's like asking a group of professional chefs what they are most worried about regarding a new, high-tech oven. The news says, "They're terrified the oven will grow legs and eat the neighborhood!" But the chefs say, "No, we're mostly worried that someone will burn the toast, the oven will leak gas, or the electricity bill will be too high."
• The Takeaway: Researchers see AI as a "normal" technology with real-world glitches, not a magical monster. They are far more concerned about misinformation, cybercrime, and people using the tools for bad purposes than about the AI gaining sentience.

2. The Optimism Gap (But Shared Fear)

The Headline: The public is often skeptical, thinking AI is mostly a scam or a danger.

The Reality: The researchers are actually quite optimistic! 87% of them believe the benefits outweigh the risks. They see AI as a super-tool that could revolutionize education and healthcare.

• The Analogy: Imagine a group of car mechanics (the researchers) and a group of commuters (the public). The commuters are terrified the new self-driving cars will crash and kill everyone. The mechanics, however, are excited because they know how to fix the brakes and believe the cars will save lives.
• The Twist: Even though the researchers are more hopeful about the good stuff, they agree with the public on the bad stuff. Both groups are equally worried about fake news, data privacy, and cyberattacks. They just disagree on how likely the "robot apocalypse" is.

3. The "Black Box" Mystery

The Headline: Everyone is worried that AI is a "black box"—a machine that makes decisions we can't understand.

The Reality: While the public is very worried about not understanding how AI thinks, the researchers are actually less worried about this than the public is.

• The Analogy: The public is like a passenger in a car who is terrified because they can't see the engine. The researchers are the mechanics; they know the engine is complex and sometimes hard to explain, but they aren't losing sleep over it because they are busy fixing the other problems (like the brakes or the fuel efficiency).

4. Who is in Charge?

The Headline: Big Tech companies are racing to build the biggest, fastest models.

The Reality: The researchers are unhappy about this. They feel that a few giant corporations are holding the steering wheel, and they want more open-source (publicly shared) models. They also want to make sure AI is used for good things like healthcare and education, not for making weapons or just making money.

• The Analogy: It's like a group of scientists who invented a new type of fertilizer. They want to share the recipe with all the farmers to help the world grow food. But a few big corporations have locked the recipe in a vault, only selling it to the highest bidder, and the scientists are frustrated that the public isn't getting the benefits.

5. The "Deficit" Myth

The Headline: Some people think the public is scared of AI because they are ignorant and don't understand science.

The Reality: The paper argues this is wrong. The researchers' worries are actually very similar to the public's worries.

• The Analogy: It's not that the public is "dumb" and doesn't get it; it's that the public is reacting to the same real problems (like job loss or privacy) that the researchers see every day. The problem is that the loud voices in the media are shouting about "super-intelligence" while ignoring the real, boring, but dangerous problems like bias in hiring algorithms or deepfakes.

The Big Conclusion

The paper ends with a warning: The conversation about AI is being hijacked by a few powerful voices. This is dangerous because it pushes the actual experts (the researchers) and the public apart.

• The Metaphor: Imagine a town hall meeting about a new bridge. Instead of listening to the engineers who built it and the citizens who will walk on it, the meeting is dominated by a few actors in costumes shouting about whether the bridge will turn into a dragon.
• The Solution: We need to stop listening to the "dragon" stories and start listening to the engineers. We need to talk about the real risks (like privacy and fairness) that both the scientists and the public care about, rather than speculating about science fiction scenarios.

In short: AI researchers are not waiting for the robots to take over. They are busy worrying about the same things you are: fake news, privacy, and making sure the technology helps everyone, not just a few rich companies. They want a real conversation, not a sci-fi movie.

Technical Summary

Here is a detailed technical summary of the paper "What are AI researchers worried about?" by O'Donovan et al. (March 2026).

1. Problem Statement

The rapid advancement of Artificial Intelligence (AI) has generated intense public debate and investment, yet the discourse is often dominated by a narrow set of voices (tech leaders and prominent scientists) focusing on extreme narratives: either utopian promises or existential threats (e.g., human extinction).

• The Gap: There is a lack of empirical data regarding the actual attitudes of the broader community of AI researchers. Previous surveys often rely on small samples, poor design, or "expert" framing that forces probabilistic predictions about catastrophic events, thereby masking the diversity and uncertainty inherent in scientific research.
• The Hypothesis: The authors hypothesize that the views of AI researchers differ significantly from the public, particularly regarding the prioritization of risks. They suspect that the public discourse overemphasizes long-term existential risks while underrepresenting immediate, sociotechnical concerns that researchers actually face.

2. Methodology

The study employs a large-scale, cross-sectional survey designed to capture the attitudes of active AI researchers globally.

• Sampling Strategy:
  • Source: Researchers were identified via the ArXiv pre-print repository (a key platform for non-academic and industry researchers), covering four subcategories: cs.AI, cs.LG, cs.CV, and cs.CL.
  • Corpus: 43,325 articles published between Jan 2020 and Feb 2024, yielding 165,226 unique authors.
  • Selection: A purposive, convenience sampling approach was used. 99,516 authors were randomly selected and invited via email.
• Data Collection:
  • Timeline: June 20 – July 22, 2024.
  • Response Rate: 7,595 surveys started; 5,318 completed. The final analytical dataset included 4,260 valid responses (7.6% response rate).
  • Demographics: 77% from academia/research orgs, 23% industry/government. 71% hold PhDs. 81% male, 18% female. Geographically diverse (USA 25%, UK 11%, China 8%, 29% EU, 17% from low/middle-income countries).
• Survey Design:
  • Comparative Framework: Questions were adapted from existing UK public surveys (Ada Lovelace/Alan Turing Institutes, ONS) to allow direct comparison between researchers and the public.
  • Key Components: Likert-scale questions on risks/benefits, agree-disagree statements on governance, and a critical open-text question: "What one thing most worries you about AI?"
• Analysis:
  • Qualitative Coding: 3,718 open-text responses were manually coded (after discarding BERTtopic modeling as ineffective). A hierarchical coding scheme was developed, resulting in 51 distinct categories.
  • Quantitative Analysis: Statistical tests (two-sample proportional z-tests) were used to identify significant correlations between demographic variables (gender, sector, AGI belief) and specific worry categories.

3. Key Contributions

• Scale: This is presented as the largest survey of AI researchers to date (n=4,260), surpassing previous studies in sample size and methodological rigor.
• Deconstruction of the "Expert" Narrative: The paper challenges the prevailing media narrative that AI researchers are primarily concerned with "existential risk" or "superintelligence." It provides empirical evidence that researchers view AI as a "normal technology" with immediate, tangible sociotechnical risks.
• Convergence vs. Divergence: It identifies a surprising convergence between researchers and the public regarding harms (disinformation, data misuse, cybercrime), while highlighting a significant divergence regarding benefits (researchers are far more optimistic).
• Reflexive Concerns: The study highlights "reflexive" concerns (issues regarding the research ecosystem itself, such as data provenance, industry control, and hype) that are often absent from public discourse.

4. Key Results

A. Risks vs. Benefits

• Optimism Gap: 87% of researchers believe benefits outweigh or balance risks, compared to only 57% of the UK public. Conversely, 36% of the public believes AI will have no positive impact, whereas <1% of researchers hold this view.
• Convergence on Harms: Researchers and the public agree on the top three negative impacts: Disinformation, Use of data without consent, and Cybercrime.
• Inequality: A vast majority of researchers agree that risks and benefits will be unequally distributed, with female researchers showing significantly higher concern for unequal impacts.

B. The "Worry" Analysis (Open-Text Responses)

The most significant finding comes from the open-text question. Only 3.4% of researchers cited "Existential Risk" (e.g., rogue AI, human extinction) as their top worry. This contrasts sharply with the 11% citing Malicious Use and 10% citing Misuse.

• Top Concerns:
  1. Malicious Use (10.6%): Intentional harm by bad actors/corporations.
  2. Misuse (9.9%): Inappropriate application of the technology.
  3. Misinformation (8.8%): Spread of false content.
  4. Impact on Jobs (7.1%): Labor displacement.
  5. Public Understanding (5.4%): Misconceptions about AI capabilities.
• Reflexive Concerns: Researchers expressed significant worry about Hype (5.1%), Bias (4.5%), Privacy (4.0%), Power Concentration (3.9%), and Profit-driven development (3.2%).

C. Governance and Responsibility

• Training Data: 65% of researchers believe there should be constraints on training models on public data (requiring explicit permission or opt-out), rejecting the "all public data is fair game" stance.
• Accountability: Two-thirds of researchers agree that creators should be held responsible for real-world impacts, rejecting the idea that responsibility lies solely with users.
• Open Source: A majority supports open-source models and believes the tech industry plays too big a role in setting research priorities.
• Funding Preferences: Researchers strongly prefer government funding for Healthcare and Education, and show low enthusiasm for Military AI.

D. Correlations and Demographics

• AGI Belief: Researchers who believe AGI is inevitable are more likely to worry about existential risk/alignment and less likely to worry about "hype." Those who doubt AGI inevitability are more concerned about hype and less concerned about privacy/alignment.
• Sector Differences: Industry researchers are more likely to claim the public worries too much about risks and are less concerned about environmental impacts compared to academic researchers.
• Gender: Women are significantly more likely to worry about Bias (8% vs 4% for men).

5. Significance and Implications

• Correcting the Discourse: The paper argues that the current public and policy discourse is distorted by a "deficit model" (assuming the public is ignorant) and a focus on speculative long-term risks. In reality, researchers are focused on immediate, distributive, and sociotechnical harms.
• Policy Recommendations: Policymakers should move away from speculating on future catastrophic risks and instead collaborate with researchers to mitigate current sociotechnical risks (e.g., copyright, labor displacement, data privacy, and algorithmic bias).
• Democratization of AI: The findings suggest a need for new forms of public dialogue that include the diverse views of the research community, rather than just the "powerful voices" of industry leaders.
• Urgency: As AI research becomes increasingly privatized, the window for constructive, open dialogue between researchers and the public may close, making this survey a critical baseline for understanding the community's values before they are potentially silenced by corporate secrecy.

Conclusion: The study concludes that AI researchers are not a monolith obsessed with "superintelligence"; they are a diverse group concerned with the "normal" risks of technology deployment, inequality, and the ethical governance of their field. Their views align more closely with public concerns about immediate harms than the "existential risk" narrative suggests.