Is the AI Safety debate facing a language barrier?
In short
Are we struggling to find a common language for AI safety? While the public is moved by visceral and "sticky" concerns like environmental impact of AI, the technical community is locked in a jargon-heavy arena where experts talk past each other and can’t seem to agree on the meaning of their own words.
Quick breakdown
Science communication is hard. Look at Climate Change: we started raising concerns in the late 50s with the first mentions of “Global Warming” in the press and the tracking of CO2 levels in the atmosphere. We’ve gone a long way since then, and yet we are still fighting to build public consensus and willingness to act.
Research has shown that the specific terms chosen to discuss climate change can influence people’s attitudes. A 2024 study by Bruine de Bruin et al.¹ found that familiar terms like "climate change" and "global warming" actually generated more concern and support for policy than "hotter" alternatives like "climate emergency" or "climate crisis", which often elicited less urgency despite their intended aim.
I wonder if we may be facing the same issue with our preparedness for AI, and I can see four underlying problems that are limiting our ability to have constructive discussions:
Problem #1, jargon overload:
There are so many terms thrown around to describe overlapping impacts, potential failures, and risks of AI. And they branch out across many domains, because AI touches everything: healthcare, education, work, cybersecurity, warfare, mental health, data privacy, arts and entertainment, government, energy, etc… The vocabulary is fragmented and definitions have not yet fully settled.
Problem #2, alarmist vs. denialist tones:
Much like the “climate crisis” label, some AI terminology is rather inflammatory. We see this with “prophets” like Eliezer Yudkowsky, whose doomsday warnings, such as "If anyone builds it, everyone dies"², undermine his own logic and call to action. On the other side of the argument, accelerationists like Marc Andreessen promote the bright future of AI and dismiss regulation as a moral crime against humanity³.
Perhaps a big difference with the “early days” of Climate Change: there are very famous, influential voices shouting about the risks and promises of AI, some of them with little technical expertise. Musk recently shared his views again⁴ about the future impact of AI on jobs, and although he is a divisive character, he still has a huge and dedicated following.
It doesn’t help that popular culture has grotesquely dramatised AI as "evil robots" in films like The Terminator or The Matrix, causing people to roll their eyes at terms like Existential Risk or Rogue AI.
Problem #3, audience literacy:
I visited family in France over the holidays. Whenever I mentioned that I study AI, everyone wanted to talk about the same thing: AI’s water consumption and job displacement. When I tried to deepen the discussion or move onto other risks, I either received dead stares or approximate quotes from TV or TikTok. It might be that I’m a terrible science communicator. Or that AI has evolved so fast in the past 3 years that we can’t expect the public to keep up with how it works and fails.
But this might not matter in the end. In the case of climate change, there was (and still is) a clear call to action for individuals. With AI, the public has far less direct agency. Outside of personal tool use or voting, there is no obvious way for a citizen to "mitigate" the risk of a frontier model. This means that although it is important to improve AI literacy in the public for their own sake, the primary target audience for AI safety is the technical community, the industry “patrons”, and the policy makers.
Problem #4, time horizon:
The conversation between experts and policy makers is often a mismatch. While engineers in frontier labs focus on the technical “how” of solving issues of the current and next model, regulators look at the “what” of long-term systemic risks. For policy to be future-proof, it cannot be too tightly linked to the shifting jargon of current AI capabilities and failures. If we write laws applying to LLMs to address "hallucinations" or "sycophancy", we risk building a regulatory framework that is obsolete by the time the next architecture arrives.
Why you should care
The claims that AI consumes a lot of water or that it could displace jobs have landed because they are visceral and relatable, unlike abstract concepts like "Agentic Misalignment". Should we make some effort to “mundanise” other AI risks into concepts that stick? Words might not change the economic system or geopolitical dynamics at play in the evolution of AI, but they could help build the “social permission” Satya Nadella argues is necessary for this technology to exist in our world⁵. We need a shared language to talk about outcomes and the real-world results we want from AI.
References
Bruine de Bruin, W., Kruke, L., Sinatra, G.M. et al. Should we change the term we use for “climate change”? Evidence from a national U.S. terminology experiment. Climatic Change 177, 129 (2024)
Eliezer Yudkowsky and Nate Soares. If Anyone Builds It, Everyone Dies. Little, Brown and Company (2025)
Marc Andreessen. The Techno-Optimist Manifesto. Andreessen Horowitz (16 Oct. 2023)
Microsoft CEO Satya Nadella: AI Boom, Energy Battle & His Complicated Alliance with Sam Altman. MD Meets podcast (29 Nov. 2025)
Bruine de Bruin, W., Rabinovich, L., Weber, K. et al. Public understanding of climate change terminology. Climatic Change 167, 37 (2021).
What’s In A Name? Global Warming vs Climate Change. Yale Program on Climate Change Communication (May 2014)