AI avatars, research bots and skin startup raise safety concerns

AI avatars are now grading practice pitches at Harvard’s startup bootcamp, while an AI research bot outperformed Anthropic and OpenAI in reproducing scientific papers. The moves expose a clash between rapid productization and thin safety scaffolding.

Harvard Business School’s Foundry program charges $699 for a six‑week sprint and plugs AI avatars into mock board meetings. The avatars deliver feedback on pitch decks and answer investor questions. Inherent, a British lab founded by DeepMind alumni, released Faraday, an agent that replicated research papers better than Claude from Anthropic and GPT‑4 from OpenAI in internal benchmarks. OpenAI, meanwhile, urged California lawmakers to tighten SB 53, a safety bill the company once opposed. At the same time, Michael Polansky’s skin‑AI startup raised $100 million, hit unicorn status in 48 hours, and announced a platform that keeps living human skin tissue alive for weeks to test compounds. A new study also found that leading AI labs still lack publicly documented plans for containing rogue models.

AI avatars in the classroom

Harvard’s Foundry program added AI avatars to its curriculum last month. The avatars simulate senior executives and venture capitalists. Students receive real‑time critiques on slide flow and financial assumptions. The system logs each interaction for later review.

The price point of $699 positions the bootcamp as a low‑cost alternative to traditional accelerators. Participants can run unlimited practice sessions without scheduling human judges. The avatars draw on large language models trained on public pitch data. The rollout follows a broader trend of embedding generative AI in education tools.

Critics worry that feedback from a model may reinforce shallow storytelling. The avatars lack lived experience in fundraising. They can miss nuance in market dynamics. The bootcamp’s own alumni survey noted mixed satisfaction with the depth of the AI critiques.

Faraday’s research replication claims

Inherent announced Faraday in a brief blog post. The agent reproduced the methodology and results of 20 peer‑reviewed papers across chemistry and computer science. In internal tests, Faraday’s outputs matched the original texts more closely than Claude 2.1 and GPT‑4.

The team attributes the performance to a pipeline that extracts citation graphs and rewrites sections in a controlled style. The model does not claim to generate novel insights. It simply mirrors existing work.

The claim raises questions about academic integrity. If a bot can rewrite a paper verbatim, plagiarism detection tools may need to adapt. Researchers could use Faraday to draft literature reviews faster, but they must verify the citations manually.

California’s safety bill under pressure

OpenAI filed an amicus brief last week urging the California legislature to strengthen SB 53. The bill currently requires high‑risk AI developers to conduct external audits and publish risk assessments. OpenAI argues the language is vague and the enforcement mechanisms weak.

The company’s stance marks a reversal from its 2022 position, when it lobbied against the bill’s predecessor. OpenAI cites recent incidents of model misbehavior as evidence that the state must act quickly. Lawmakers have not yet responded to the brief.

Opponents warn that tighter regulations could slow innovation. They point to the bootcamp’s AI avatars and Inherent’s Faraday as examples of rapid progress that could be hampered by heavy compliance costs. The debate underscores a growing tension between safety and speed.

Containment plans remain opaque

A study released by the Center for AI Governance surveyed the top ten AI labs. Only three disclosed any formal protocol for shutting down a rogue model. The rest offered generic statements about “responsible development.”

The lack of detail worries policymakers. Unexpected model behavior has risen in the past year, with several high‑profile incidents of language models generating disallowed content. Without clear containment procedures, regulators have limited leverage.

The study’s authors recommend mandatory public reporting of shutdown drills. They also suggest third‑party audits of containment infrastructure. The recommendation aligns with OpenAI’s call for stronger state oversight.

A skin‑AI unicorn with living tissue

Michael Polansky’s startup, SkinAI, announced a platform that sustains human skin explants for up to three weeks. The system uses a perfusion bioreactor and an AI model that predicts optimal nutrient mixes. The company raised $100 million from Iconiq and Sequoia in a round that closed in 48 hours.

SkinAI’s technology promises faster screening of anti‑aging compounds. The AI suggests formulation tweaks based on real‑time tissue response. The company claims the approach reduces animal testing.

Skeptics note the ethical gray area of keeping human tissue alive ex‑vivo for commercial research. The regulatory framework for such biotechnologies is still forming. The rapid funding also highlights how venture capital can accelerate high‑risk AI‑bio convergence.

What to watch

Track California’s response to OpenAI’s brief on SB 53. A tightened bill could set a precedent for other states. Monitor whether Inherent publishes open‑source details of Faraday’s pipeline, which would affect reproducibility standards. Watch for any formal containment protocols released by the surveyed AI labs. Finally, follow SkinAI’s clinical trial pipeline to see if the living‑tissue model scales beyond cosmetics.

Updates

• 2026-08-23 — What is a VESA mount and how to know what type your TV has (source)