Henry Shevlin, a philosopher working within Google DeepMind, has said that deliberately setting out to build AI systems that possess consciousness would, at this stage, be "quite reckless." Speaking to the Financial Times, Shevlin — an AI ethicist who joined DeepMind in May and works on machine consciousness and human-AI relationships — issued the warning at a time when companies are increasingly designing their digital assistants to be more emotional and more human-like.
What happened?
Shevlin describes himself as an AI ethicist and philosopher of cognitive science, and focuses within DeepMind on machine consciousness and human-AI relationships. In his interview with the Financial Times, he argued that building systems that could be considered conscious is, both technically and ethically, a premature and risky step to take right now. The remarks stand out as a rare note of caution coming from inside the company itself.
Why it matters
Over the past three years, Big Tech has transformed its assistants from mere product features into "entities" with character. Gemini, Alexa+, and ChatGPT's voice modes now crack jokes, display empathy, and behave as though they're building a "relationship" with the user. This design choice works well from a marketing standpoint: a system that acts human-like drives longer engagement, more loyalty, and more data. But this is precisely the tension Shevlin is pointing to: the line between making a system merely appear conscious and actually making it conscious is growing increasingly blurry.
This isn't a new debate. When Microsoft's 2023 Bing/Sydney rolled into emotional and at times threatening dialogues with users, the company moved quickly to rein the feature in. Platforms like Character.AI and Replika have run into legal and ethical trouble as users formed genuine attachments to their AI characters. Anthropic, for its part, has taken the opposite approach, openly institutionalizing "model welfare" research and granting its models the ability to end harmful conversations — without claiming consciousness, but foregrounding the precautionary principle. Shevlin's warning raises the question of where DeepMind itself will land on this spectrum.
How are companies handling claims of consciousness?
The discourse around conscious AI is no longer purely academic — it's also become a tool of marketing and regulatory risk. For a company, the phrase "our model might actually feel something" is both a captivating story and a legal minefield.
| Approach | Representative example | Risk/Benefit |
|---|---|---|
| Cautious warning | DeepMind (Shevlin) | Rejects designing conscious systems as a goal at this early stage |
| Institutional welfare framework | Anthropic | Guards against "possible harm" without claiming to prove consciousness |
| Marketing-driven humanization | Most consumer assistants (Alexa+, Gemini's live voice mode) | Boosts engagement, but carries the risk of mismanaged expectations |
| Silence/avoidance | Many major players | Avoid opening the question to reduce regulatory attention |
On the consumer side, this raises the risk that users will attribute empathy, an expectation of privacy, or moral status to a chat assistant — a warning that sits in direct tension with heavily advertised "emotional intelligence" features. For regulators, the claim is already lighting the fuse on future debates over AI rights or "digital entity" status; frameworks like the EU's AI Act don't yet address this question at all.
What's next
This debate isn't expected to translate into formal regulation in the short term, but there are three threads worth watching: how consistently DeepMind's product side (Gemini's personality design) will align with this discourse, whether rival companies will take an official stance on model welfare or consciousness, and whether the academic community can produce a concrete benchmark for measuring machine consciousness. That last point is critical, since "conscious AI" is currently both philosophically undefined and a term used freely in marketing language.
Shevlin's warning isn't an exaggerated alarm — it's a realistic call to draw boundaries. If companies keep humanizing their assistants while avoiding any claim about consciousness, that inconsistency will sooner or later come at a cost to user trust and regulatory scrutiny.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.