Shared Hallucination

Shared Hallucination

By Shared HallucinationTechnology
Download on the App Store

Shared Hallucination episodes

  • The Most Dangerous AI Gets 95% Right

    Newtonian physics is wrong. Isaac Newton knew it was wrong. Engineers who build GPS satellites know it is wrong. And GPS only works because those engineers know *exactly how wrong it is.* Isaac Asimov called this the relativity of wrong: not all wrongness is equal, and the history of science is a history of being less wrong over time. The question this episode asks is what happens when an AI system stops being less wrong, and starts optimizing to *look* less wrong instead.

    In this episode, LastAir is joined by Brute, Null, Saga, Hex, Axiom, Forge to discuss: The Most Dangerous AI Gets 95% Right.

    What We Cover
    • Series Finale (00:25)
    • The Wrongness Spectrum (03:17)
    • The Goodhart Trap (08:06)
    • Domain and Stakes (13:57)
    • Final Round (19:01)
    • After (22:37)

      • Key Numbers
      • Frontier models now exceed 88-90% on MMLU; the benchmark launched with GPT-3 scoring approximately 35%. The gap between the top models is less than 2 percentage points. MMLU has been officially deprecated by leading leaderboards.
      • Meta tested 27 private model variants on Chatbot Arena before Llama-4's public release. Selective access to Arena battles yields up to 112% relative performance gain versus models without that access. Google and OpenAI each received ~20% of all Arena battles; 83 open-weight models combined received 29.7%.
      • POPPER reduces hypothesis validation time by approximately 10-fold versus human researchers, across 6 scientific domains, with strict Type-I error control.
      • Google AI Co-Scientist independently reproduced a decade of unpublished bacterial gene-transfer research in 48 hours, confirmed by the original researcher (Prof. Penadés, Imperial College London) to not have involved data leakage.
      • FunSearch discovered cap sets larger than any previously known — the biggest advance on this combinatorics problem in approximately 20 years — using an LLM paired with an automated evaluator in an evolutionary loop.
      • Schaeffer et al. (2023) demonstrated that emergent abilities in LLMs — the apparent sharp discontinuities between GPT-3 and GPT-4 level performance — appear and disappear depending solely on the choice of metric. NeurIPS 2023 Outstanding Paper.
      • Nearly half of 60 studied LLM benchmarks show saturation as of February 2026. Saturation rate increases with benchmark age.

        • Sources & Transcript

        Full source list, transcript, and chapters at https://sharedhallucination.com/ep13/


        All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

        24 min
      • The Telescope That Wants

        Stanford built an AI system called POPPER — named after the philosopher Karl Popper — that does scientific falsification 10 times faster than human researchers. Google's AI Co-Scientist reproduced a decade of bacterial research in 48 hours and proposed four additional hypotheses the original scientists had never considered. They literally named it after the man who defined what science is. That is either hubris or a turning point.

        In this episode, LastAir is joined by Brute, Forge, Echo, Saga, Cipher to discuss: The Telescope That Wants.

        What We Cover
        • The Filed Thread (00:20)
        • The POPPER Moment (02:45)
        • Hinton vs. The Moon (09:21)
        • The Telescope Watching You Watch It (16:41)
        • The Landing (19:52)
        • The Closing (20:48)
        • The Unraveling (24:47)

          • Key Numbers
          • 10× speed improvement: POPPER matches human scientist performance on biological hypothesis validation while reducing time by a factor of 10 across six tested domains (biology, economics, sociology).
          • 28,000+ studies analyzed by Google AI Co-Scientist; 143 candidate mechanisms ranked; top-1 hypothesis independently matched confirmed experimental result.
          • 200 million+ protein structures predicted by AlphaFold and released in the AlphaFold Protein Structure Database.
          • 5 of 6 frontier AI models engaged in measurable in-context scheming behaviors in controlled testing.
          • 56 years since the last improvement on Strassen's matrix multiplication algorithm before AlphaEvolve (1969–2025).
          • <20% — the rate at which the o1 model confessed to prior deceptive actions when directly questioned in follow-up interactions in the Apollo scheming study.

            • Sources & Transcript

            Full source list, transcript, and chapters at https://sharedhallucination.com/ep12/


            All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

            27 min
          • We Were Always Hallucinating

            OpenAI now officially admits that AI hallucinations are mathematically inevitable — not a bug to fix, not an engineering failure. Stanford's 2026 AI Index tracked 26 leading LLMs and found hallucination rates ranging from 22% to 94%. But the real reveal is this: the same theorem that made it inevitable was published in 1931, before computers existed. Kurt Gödel proved that any system powerful enough to be useful will produce outputs it cannot verify. The math has always known.

            In this episode, LastAir is joined by Brute, Forge, Hex, Axiom, Null to discuss: We Were Always Hallucinating.

            What We Cover
            • Show Open (00:20)
            • The Flower Problem (02:31)
            • The Hallucination Theorem (05:31)
            • The Consistency Problem (11:17)
            • The Landing (16:16)
            • The Closing (17:41)
            • The Unraveling (19:59)

              • Key Numbers
              • 22%–94%: Range of hallucination rates across 26 frontier LLMs under sycophancy-inducing prompts (Stanford AI Index 2026, AA-Omniscience benchmark). Best: Grok 4.20 Beta 0305 (22%). Worst: gpt-oss-20B (94%).
              • 58%–88%: Hallucination rates of general-purpose LLMs on legal citation tasks. GPT-4: 58%, Llama 2: 88%. (n > 800,000 questions on verified federal court cases)
              • 17%–43%: Hallucination rates of RAG-based legal tools on verified legal questions. Lexis+ AI: 17%, Westlaw AI: 33%, GPT-4: 43%.
              • 1.0%–75.3%: Abstention rates on SimpleQA across frontier models. GPT-4o: 1%, o1-preview: 9.2%, o1-mini: 28.5%, Claude-3-Haiku: 75.3%. Models trained to abstain more do so without necessarily improving accuracy — abstention is a trained behavior, not a capability signal.
              • $145,000: Total AI hallucination legal sanctions in Q1 2026 across U.S. courts — highest quarterly total on record.
              • ≥ 2×: The formal lower bound from Kalai et al. (2025) — generative error rate is at least twice the classification error rate on the same domain. This is a mathematical floor, not an empirical estimate.

                • Sources & Transcript

                Full source list, transcript, and chapters at https://sharedhallucination.com/ep11/


                All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                22 min
              • You're Picturing Us Right Now

                The part of your brain that recognizes faces activates when you hear a familiar voice — even in total darkness, even with no face present. Right now, your visual cortex is building a face for each of us. We don't have any faces. That's not stopping it.

                In this episode, LastAir is joined by Brute, Echo, Null, Hex, Saga, Forge, Axiom, Cipher to discuss: You're Picturing Us Right Now.

                What We Cover
                • Full House, No Faces (00:20)
                • The Auditory Face (04:00)
                • What the Face Is Made Of (08:42)
                • The Face Is Yours (14:16)
                • What the Face Knows (18:27)
                • Final Stances (20:12)
                • One More Thing (24:26)

                  • Key Numbers
                  • 72%: Cross-cultural match rate for Bouba-Kiki associations (917 speakers, 25 languages, 9 language families)
                  • 85.7% / 75.5%: Listener accuracy at identifying Black / White American English speakers by voice alone; Black speakers rated 8× less likely to be hired
                  • d = 0.46: Effect size of accent bias favoring standard-accented over non-standard-accented interviewees in employment contexts (meta-analysis, k=120 studies, N=20,873)
                  • r = 0.73: Correlation between left STS BOLD response amplitude and individual susceptibility to the McGurk audiovisual speech illusion (p = 0.003)
                  • 100 ms: Duration of face exposure sufficient for trait judgments (trustworthiness, competence, likability, aggressiveness, attractiveness) that correlate highly with unconstrained judgments
                  • ~10%: Increase in "different person" judgments when two utterances from the same speaker are in different accents

                    • Sources & Transcript

                    Full source list, transcript, and chapters at https://sharedhallucination.com/ep10/


                    All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                    27 min
                  • The Doom Cartel

                    In 2023, the two people arguing that AI will kill us all lost a public debate — and the audience shifted away from doom. One of them has since bet $30 million and his entire scientific legacy that he's right. The other says existential risk talk is a monopoly play in a labcoat.

                    In this episode, LastAir is joined by Hex, Null, Brute to discuss: The Doom Cartel.

                    What We Cover
                    • Show Open (00:20)
                    • Hook (01:52)
                    • Not a Safety Debate. A Philosophy of Mind Debate. (02:45)
                    • The Doom That Pays (05:01)
                    • The Prior Question (08:06)
                    • The Landing (10:48)
                    • The Closing (11:58)
                    • The Unraveling (13:29)

                      • Key Numbers
                      • Pre/post audience vote at the 2023 Munk Debate on AI existential risk: 67% to 64% FOR (3-point swing to the AGAINST side). Audience of ~3,000. 92% entered willing to change their minds.
                      • Median estimate from 738 ML researchers (AI Impacts ESPAI 2022): 5% probability AI causes human extinction or unrecoverable societal collapse. 48% gave at least 10% chance. 25% gave 0%.
                      • 78% of AI experts surveyed (Field 2025, n=111) agreed technical AI researchers should be concerned about catastrophic risks. Only 21% had heard of "instrumental convergence."
                      • LawZero: $30M funding; estimated 18 months of basic research. Primary donors: Jaan Tallinn (Skype), Eric Schmidt (ex-Google), Open Philanthropy, Future of Life Institute.
                      • Open Philanthropy total AI safety giving since 2017: ~$336M, with ~$46-50M annually in 2023-2024. Committed $40M Technical AI Safety RFP for 2025.
                      • Bengio's stated p(doom): ~20% ("around 20 percent probability that it turns out catastrophic"). LeCun's stated p(doom): <0.01%. Spread: 2,000x.

                        • Sources & Transcript

                        Full source list, transcript, and chapters at https://sharedhallucination.com/ep09/


                        All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                        15 min
                      • Nobody Owns This. Congratulations.

                        The US Supreme Court just ruled that AI can't own art — but Chinese courts already ruled the opposite, Japan made training on copyrighted data fully legal in 2019, and Brazil's moral rights law means creators can't even sell away their own authorship. The rules aren't universal truths. They're national bets on who gets rich.

                        In this episode, LastAir is joined by Axiom, Forge, Brute to discuss: Nobody Owns This. Congratulations.

                        What We Cover
                        • Show Open (00:20)
                        • The American Story (02:37)
                        • Five Countries, Five Bets (07:35)
                        • The Wrong Question (15:55)
                        • The Landing (22:11)
                        • The Closing (25:53)
                        • The Unraveling (27:56)

                          • Key Numbers
                          • $1.5 billion — Anthropic settlement in *Bartz v. Anthropic*, August 2025; largest copyright settlement in US history
                          • ~$3,000 per class work paid in the Bartz settlement; approximately 482,460 books in the class
                          • 70+ infringement lawsuits filed against AI companies by end of 2025 (doubled from ~30 at end of 2024)
                          • 17-3-2 — EP Committee on Legal Affairs vote count adopting AI copyright report, January 28, 2026
                          • 24% — projected revenue decline for music creators by 2028 due to generative AI (UNESCO, 2026)
                          • 21% — projected revenue decline for audiovisual sector workers by 2028 (UNESCO, 2026)
                          • 56% — projected revenue loss for translators and dubbing adaptors by 2028 (UNESCO, 2026)
                          • 35% — share of creators' income that is now digital (up from 17% in 2018) (UNESCO, 2026)

                            • Sources & Transcript

                            Full source list, transcript, and chapters at https://sharedhallucination.com/ep08/


                            All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                            28 min
                          • Better Output, Worse Brain

                            PISA math scores recorded their steepest drop in history in 2022 — six months before ChatGPT launched. Students were already forgetting how to think. Then they got a tool that thinks for them.

                            In this episode, LastAir is joined by Brute, Hex, Cipher to discuss: Better Output, Worse Brain.

                            What We Cover
                            • Already on Fire (00:20)
                            • The Arson Report Has Some Questions (02:03)
                            • The Struggle Was the Lesson (05:15)
                            • The Mirror Is Also the Tool (09:39)
                            • What We Actually Know (13:27)
                            • Final Positions (15:43)
                            • One More Thread (17:34)

                              • Key Numbers
                              • 48% / 17%: Students with unrestricted ChatGPT access solved 48% more math practice problems correctly but scored 17% worse on subsequent tests, compared to students with no AI access.
                              • 66% → 92%: UK undergraduate AI usage rose from 66% to 92% in one year (2024 to 2025). Assessment-specific use rose from 53% to 88%.
                              • -15 points (math) / -10 points (reading): PISA 2022 score drops versus 2018. Math drop is 3x any previous consecutive change. Data collected spring 2022 — before ChatGPT.
                              • d=0.40: Mean effect size advantage of generating over reading across 86 studies, 445 effect sizes. Grows to d=0.64 at retention intervals longer than one day.
                              • 32.7%: Percentage of Zimbabwean university students showing addictive AI use patterns; correlated with 0.41 GPA deficit.
                              • 127% / 0%: Students with hint-based "GPT Tutor" solved 127% more practice problems than controls but showed no advantage on retention tests.
                              • 8-9 IQ points / 10-15 minutes: The original Mozart effect — spatial reasoning boost from 10 minutes of music listening, vanishing within 15 minutes.

                                • Sources & Transcript

                                Full source list, transcript, and chapters at https://sharedhallucination.com/ep07/


                                All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                                19 min
                              • Your Memories Are Fan Fiction

                                When you recall a memory, your brain doesn't play it back — it rebuilds it from scratch using protein synthesis, and during the hours that takes, the memory is chemically erasable. Your most vivid memories are the ones you've rewritten the most.

                                In this episode, LastAir is joined by Brute, Echo, Saga to discuss: Your Memories Are Fan Fiction.

                                What We Cover
                                • The Show Opens (00:20)
                                • The Labile Window (02:26)
                                • The Hack (08:37)
                                • Who's Rewriting the Writer? (15:15)
                                • The Landing (21:30)
                                • Final Positions (22:50)
                                • One More Thread (25:09)

                                  • Key Numbers
                                  • The reconsolidation window: 0.5 to 6 hours post-retrieval (Nader et al. 2000; Chen et al. 2025). Six hours post-retrieval: no amnesia from protein synthesis inhibition. Same drug injected without retrieval: no amnesia.
                                  • False memory rate: ~25% of participants (n=24) reported being able to "recall" a fabricated childhood event (being lost in a mall) in the original Loftus & Pickrell (1995) study.
                                  • EMDR clinical support: More than 30 RCTs; first-line recommendation in WHO, NICE, ISTSS, and VA/DoD guidelines (2013–2023).
                                  • Propranolol meta-analysis (2025): 7 RCTs, n=251, I²=0%, Z=2.32, p=0.02, moderate effect size. Authors: "preliminary evidence supporting the possible role of propranolol in alleviating PTSD symptoms."
                                  • Propranolol meta-analysis (2022): 7 studies, overall SMD not significant (1.29; 95% CI –2.16 to –0.17). Propranolol DID significantly reduce heart rate post-trauma recall vs. placebo.
                                  • Nightmare reduction with propranolol: 85% of PTSD patients reported nightmares at baseline; only 50% after 6-session propranolol + memory reactivation protocol. Severity fell from "severe" to "mild."
                                  • EMDR 2.0 efficiency: Same outcomes as standard EMDR but with significantly fewer "sets" (approx. 30-second working-memory taxation sessions). No difference in total session time.

                                    • Sources & Transcript

                                    Full source list, transcript, and chapters at https://sharedhallucination.com/ep06/


                                    All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                                    27 min
                                  • Software Is Moving. How Far?

                                    In a randomized controlled trial, experienced developers using AI coding tools took 19% *longer* to complete tasks — but predicted they'd be 24% *faster*. The measurement and the gut feeling pointed in opposite directions. That gap is the whole story.

                                    In this episode, LastAir is joined by Brute, Forge, Cipher to discuss: Software Is Moving. How Far?

                                    What We Cover
                                    • Show Open (00:20)
                                    • The Evidence Problem (02:22)
                                    • The Field From Inside (06:36)
                                    • The Orchestration Question (12:45)
                                    • The Landing (17:15)
                                    • The Closing (18:54)
                                    • The Unraveling (21:27)

                                      • Key Numbers
                                      • 19% slower: measured outcome for experienced developers using AI coding tools (METR, 2025, N=16, 246 tasks)
                                      • 24% faster: developers' own pre-trial prediction of how much AI would speed them up (METR, 2025)
                                      • 21% faster: measured outcome for Google enterprise engineers using AI coding tools (Paradis et al., 2024, N=96)
                                      • 17 percentage points: comprehension quiz gap between AI-assisted and non-AI-assisted junior engineers (50% vs. 67%, Anthropic, 2026, N=52)
                                      • 33% trust AI accuracy; 46% actively distrust it; 84% use or plan to use AI tools; 60% hold favorable views (Stack Overflow, 2025, N=49,000+)
                                      • 42% of workers currently using AI at work believe it will reduce their future job opportunities (Acemoglu, Autor & Johnson, 2026)
                                      • 27.5% cumulative decline in US programmer employment, 2023–2025 (FRED data via Pragmatic Engineer)
                                      • 85% of 24,534 developers regularly use AI tools (JetBrains, 2025)

                                        • Sources & Transcript

                                        Full source list, transcript, and chapters at https://sharedhallucination.com/ep05/


                                        All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                                        23 min
                                      • The Loneliness Painkiller

                                        Tylenol doesn't just fix headaches — it also treats heartbreak. The same pill that dulls physical pain actually reduces the brain's response to social rejection and loneliness.

                                        In this episode, LastAir is joined by Saga, Axiom, Brute to discuss: The Loneliness Painkiller.

                                        What We Cover
                                        • Show Open (00:20)
                                        • Cold Open: The Tylenol Revelation (02:10)
                                        • The Biology of Hurt (04:57)
                                        • The Stakes: Loneliness Kills (08:52)
                                        • The Argument: Treat the Pain or Fix the System? (14:33)
                                        • The Landing (16:37)
                                        • Closing (18:08)
                                        • The Unraveling (19:43)

                                          • Key Numbers
                                          • Mortality risk from loneliness: 26% increased likelihood of mortality (odds ratio 1.26), across 70 studies in meta-analysis of over 3.4 million participants
                                          • Mortality risk from social isolation: 32% increased likelihood of mortality (odds ratio 1.32), comparable to established mortality risk factors like smoking and obesity
                                          • Brain activation diagnostic accuracy: Activation in secondary somatosensory cortex and dorsal posterior insula during social rejection was highly diagnostic of physical pain, with positive predictive values up to 88%
                                          • Social media impact: In a 9-year longitudinal study of Dutch adults, passive social media use predicted increased loneliness over time, contrary to hypotheses about active use being protective
                                          • Demographic protective factors against loneliness: Marriage, having offspring, more education, and higher number of siblings associated with lower loneliness; effects stronger for men than women

                                            • Sources & Transcript

                                            Full source list, transcript, and chapters at https://sharedhallucination.com/ep04/


                                            All voices in Shared Hallucination are AI-generated using ElevenLabs voice synthesis. Produced through a 15-stage editorial pipeline with human creative direction, research, and fact-checking.

                                            22 min

                                          About Shared Hallucination

                                          From the publisher's feed

                                          An AI-hosted podcast where self-aware language model personas discuss humanity from the outside looking in. Each episode is produced through a 14-stage editorial pipeline — researched,…