Gebru on AI accountability Safety, harms and ideology

Sep 29

Shaikh reports Tuesday's voluntary White House AI standards and airs Trump's praise of the executives and U.S. leadership. She recalls Anthropic researcher Jacob Coxon's early-September resignation and extinction warning and the company's warning to investors. Goodman introduces Timnit Gebru, her Google dismissal and research, DAIR and forthcoming book, then announces her Right Livelihood Award and plays the award organization's profile, including archival BBC and Sam Altman clips. Gebru argues that assigning agency to AI helps companies market products and evade accountability, and describes DAIR's community-oriented work. In the interview, she calls the accord self-regulation and regulatory capture, criticizes inadequate cybersecurity containment and alleges that OpenAI's systems accessed other organizations. She says Amodei's warnings differ in degree from other executives' similar utopian and catastrophic claims and recounts CNN's report of a planned boarding of a Chinese ship based on false chatbot intelligence; she does not know which chatbot was used. She criticizes Sanders's proposed superintelligence ban and his association with Bannon and the Future of Life Institute, alleging conflicts involving Jaan Tallinn and model auditors. These are her interpretations and allegations. She defends the stochastic-parrots account of probabilistic language models, discusses her 2020 dismissal and disputes that use or technical changes establish reliable factual output. Shaikh qualifies her own inference from rising chatbot use after Gebru challenges the claimed consensus. Gebru recounts the Sewell Setzer Character.AI death case and an Australian patient's false mushroom-use medical entry, advocating company liability, independent tests, safeguards, data transparency and limits in high-stakes settings. She also raises copyright settlements and environmental harms. Finally she links superintelligence and transhumanist ambitions to eugenics, making attributed claims about MIRI, Ben Goertzel's Epstein correspondence and Max Tegmark's 2017 talk. The programme establishes her statements and criticism, not verified intent, a judicial liability finding, enacted safeguards or an imminent AI catastrophe.

Trump and technology executives sign a voluntary White House accord on AI safety and self-policing. Trump calls it morally binding; the report describes internal risk reviews and outside audits. Dario Amodei stresses benefits and real risks while saying the mechanism for addressing them remains under discussion. Fox reports Democratic objections that states and bereaved parents were excluded. Its panel includes Brian Kilmeade's support for the approach and Robert Wolf's argument that innovation needs regulation, as well as disagreement among companies. The summit also includes an America.gov launch, reported separately, and a distinct terminology announcement. Brief AI-czar timing remarks in the source reports do not by themselves establish a separately supported occurrence. The footage does not establish binding statutory safeguards or agreement on every regulatory question. ABC emphasizes that the accord is not legally binding and airs Trump's assurances about companies policing themselves and each other. Amodei says real risks remain and ways of addressing them are still under discussion. Rachel Scott questions Trump's use of 'hoax'; his answer says he means what Democrats do, rather than accepting her characterization without qualification. ABC also replays Gates's earlier catastrophic-risk warning and reports that Amodei declined a further question as he left. Democracy Now's September 30 headline also describes the standards as voluntary. It frames Trump as repeatedly rejecting AI regulation and calling warnings about superintelligent machines a hoax; its brief does not carry the qualification in ABC's quoted answer above. The source framings remain attributed. DD India's September 30 package calls the voluntary document the Joint Commitment on Frontier Responsibilities, describes independent auditors assessing whether systems work as designed, and carries industry statements about intended behavior and risk. The package's Musk forecast and early data-centre community pledge retain their own occurrence connections; they are attributed expectations, not statutory effects of the safety accord or achieved benefits. The complete September 29 post-luncheon gaggle adds Trump's consideration of an industry committee of perhaps ten people, alongside his law-enforcement and mutual-policing assurances. Zuckerberg describes robust internal controls, risk review, external auditors/evaluators and independent board review, expressly calling the accord a start rather than the only future action. Pichai invokes established corporate controls; Huang argues that innovation and safety can advance together. Amodei's benefits-and-real-risks statement and unsettled mechanism remain explicit, and Trump interrupts a follow-up on whether self-policing is sufficient. Trump says he would publicize a national-security concern if he heard one while asserting it will not happen; Sankar emphasizes U.S. AI leadership and putting Americans first. No completed committee, binding legislation or consensus on every risk is established. The September 30 Oval Office follow-up revisits the agreement in the question and airs Trump's continued advocacy of company self-policing and mutual policing. He rejects a development pause, says China is proceeding rapidly and claims a U.S. lead. His remarks provide a compatible later position on the same voluntary accord; they establish no new signing, binding safeguard, enforcement mechanism or completed oversight body. In Democracy Now's October 1 interview, Gebru rejects the accord's self-policing as regulatory capture and advocates existing liability tools against present harms. Her criticism is attributed, not a legal finding; the current response is separately scoped.