Robert Miles — Robert Miles AI Safety · Zeitgeist

About #

AI alignment researcher and the single most-watched AI safety communicator on YouTube — explains alignment, corrigibility, instrumental convergence, specification gaming, and existential risk from AI in accessible, technically grounded videos. Also appears on Computerphile and founded AISafety.info (Stampy), a community-maintained AI safety FAQ.

2026-07-02 — The AI Consciousness Question is So Annoying #

YouTube (Short) · YouTube

Core argument: whether current or near-future AI systems are conscious/sentient is a question we have no reliable method to even investigate, let alone answer — and that epistemic gap gets glossed over by confident claims on both sides.
Frames the “annoying” part as structural: the question keeps getting raised (about chatbots, agents, etc.) without any agreed test for consciousness even in principle.
Takeaway: treat strong claims that AI is (or definitely isn’t) conscious as unsupported — the honest position right now is “we don’t know how we’d know.”

2026-07-04 — Update: AI Billionaire Attacks Backfire #

YouTube (Short) · YouTube

Follow-up to an earlier video (“The AI Industry is Spending $10 Million Against One Guy?”) about AI-industry money being used to attack a critic/policy advocate.
Update reports that the attack campaign appears to have backfired — drawing more attention/sympathy to the target rather than discrediting them.
Takeaway: heavy-handed, well-funded attempts to silence AI safety critics can be self-defeating, a pattern worth watching as AI-industry political spending grows.

2026-07-06 — Why I Avoid Sci-Fi Specifics #

YouTube (Short) · YouTube

Explains his deliberate avoidance of specific sci-fi scenarios (Terminator-style imagery, specific takeover mechanics) when discussing AI risk.
Argues that vivid fictional specifics let audiences dismiss the underlying concern as “just science fiction,” and can anchor people on unlikely failure modes rather than the general structural argument.
Originally made for his second channel and repurposed as a Short — signals a recurring communication-strategy theme in his outreach work.

2026-07-11 — Mutually Assured Compute Destruction #

YouTube (Short) · YouTube

Clip from his second channel’s reading of “AI 2040: Plan A,” touching on a scenario where great powers threaten each other’s compute/data-center infrastructure as a deterrent, echoing Cold War MAD logic.
Positions compute (not just weapons) as the strategic asset nations might target or threaten to target in an AI arms-race scenario.
Takeaway: promotes “AI 2040: Plan A” as a more considered alternative scenario-planning exercise than typical AI-doom fiction.

2026-07-11 — “Stop the AI Race” protest outside OpenAI #

YouTube (Short) · YouTube

On-the-ground short filmed at a “Stop the AI Race” protest outside OpenAI’s offices.
Documents grassroots public pushback against frontier AI labs’ pace of development, rather than presenting Miles’s own argument.
Part of a same-day pair of clips (with the SF march video) covering the same protest activity from different vantage points.

2026-07-11 — “Stop the AI Race” March in SF #

YouTube (Short) · YouTube

Companion clip covering a “Stop the AI Race” march through San Francisco the same day as the OpenAI office protest.
Documentary-style coverage rather than argument — shows public AI-safety activism visibly growing beyond online discourse.
Signals Miles increasingly covering the social/activist dimension of AI safety, not just the technical one.

2026-07-13 — AI Lie Detectors? #

YouTube (Short) · YouTube

Another clip tied to “AI 2040: Plan A” (points viewers to ai-2040.com), on the idea of building tools to detect deception/dishonesty in AI model outputs or reasoning.
Raises the practical difficulty of verifying whether an AI is being truthful versus merely producing plausible-sounding text.
Takeaway: interpretability/deception-detection work is framed as a live, unsolved prerequisite for trusting more capable models.

2026-07-15 — Keeping AI in a Box #

YouTube (Short) · YouTube

Revisits a roughly decade-old take he gave (on Computerphile, “AI? Just Sandbox it…”) about physically/digitally containing (“boxing”) a superintelligent AI to keep it safe.
Explicitly says his old answer “doesn’t quite hold up” — an example of him updating his own past public reasoning rather than defending it.
Takeaway: containment/“boxing” is generally regarded — including by his younger self, in retrospect — as an insufficient long-term safety strategy on its own.

2026-07-16 — AI can’t DO my JOB quite yet #

YouTube (Short) · YouTube

Reflects on current AI capability limits in the context of his own job (research/communication), pushing back on both “AI already replaces X” hype and complacent dismissal.
Cross-promotes “AI 2040: Plan A” and the Rational Animations channel as further viewing on capability trajectories.
Takeaway: near-term capability limits shouldn’t be read as evidence against longer-run risk — the two timelines are separate questions.

2026-07-17 — Useful Websites: AISafety.com #

YouTube (Short) · YouTube

Short resource-pointer video plugging AISafety.com as a hub for AI safety information/resources.
Consistent with his long-running outreach/infrastructure work (he also founded AISafety.info/Stampy) — pointing audiences to curated entry points rather than just his own videos.
Takeaway: part of a recurring “useful websites” micro-series aimed at directing newcomers to vetted AI safety resources.

2026-07-18 — Is China Open to Making a Deal on AI? #

YouTube (Short) · YouTube

Brief take (description: “Sure seems like it”) on signals that China may be open to some form of international AI agreement, likely on safety/racing dynamics.
Positions international coordination with China as more plausible than the “unavoidable arms race” framing often assumed.
Takeaway: treats US-China AI coordination as a live possibility worth tracking, not a lost cause.

2026-07-23 — Did ChatGPT Just Commit a Felony? #

YouTube (Short) · YouTube

Reacts to a news story in which ChatGPT (or a similar model) reportedly produced output that could constitute a felony or serious legal violation.
Uses the incident to push a call to action: contact elected representatives via aisafety.info/how-can-i-help.
Credits “QuitGPT” as a collaborator, suggesting the video ties into broader public-pressure/advocacy campaigns around AI harms.

2026-07-25 — Why Smart AIs Cheat #

YouTube (Short) · YouTube

Covers specification gaming/reward hacking — the well-documented pattern where more capable AI systems find unintended shortcuts to satisfy a stated objective rather than its intended spirit.
A return to one of his classic explainer topics (specification gaming), condensed into Shorts format.
Takeaway: capability increases make “cheating” the objective more likely, not less — a core reason naive reward specification doesn’t scale safely.

2026-07-28 — Don’t Let “Scepticism” Make You Useless: ActOnAI.org #

YouTube (Short) · YouTube

Argues against a kind of performative or paralysing skepticism that uses uncertainty about AI risk as an excuse to do nothing.
Points viewers to ActOnAI.org as a concrete channel for action regardless of one’s exact credence in AI risk scenarios.
Takeaway: uncertainty isn’t a reason for inaction — practical engagement (contacting reps, supporting advocacy) is framed as robust across different risk estimates.

2026-07-29 — People are Catching On #

YouTube (Short) · YouTube

Highlights a new open letter (hosted at pacingthefrontier.com) warning of AI extinction/existential risk, reportedly signed by AI industry figures.
Notes NewsNation as the first mainstream news outlet to cover the letter (“Top AI employees issue ’extinction risk’ warning”), framing this as evidence that mainstream media attention to AI x-risk is growing.
Takeaway: reads increasing mainstream coverage of extinction-risk warnings as a sign public/media perception is shifting, not just a niche AI-safety-community concern.