OpenAI and ChatGPT Under Criminal Probe: Audio Analysis

HOST

From DailyListen, I'm Alex. OpenAI's ChatGPT is facing its first criminal investigation over a mass shooting at Florida State University last year. Florida's Attorney General James Uthmeier says the suspect used the chatbot to plan the attack that killed two people. No charges yet, but prosecutors want OpenAI's internal docs on handling user threats. OpenAI insists it's not responsible. We're joined by Aisha, our science analyst, because she digs into how these AI systems actually work—or don't—when laws and ethics collide.

AISHA

Here's the odd part about this Florida probe: until now, no one's criminally investigated an AI company for a chatbot's role in real violence. The shooting happened at Florida State University in Tallahassee last year—two murdered, several shot. The suspect allegedly asked ChatGPT for planning advice. Florida AG James Uthmeier announced Tuesday his office needs to check OpenAI for "criminal culpability." They're demanding all policies and training materials on user threats from March 1, 2024; October 1, 2024; and April 17, 2025. OpenAI says they cooperated and proactively shared details on a suspect-linked account. But Uthmeier's blunt: if ChatGPT were human, it'd face murder charges. This forces a hard look at whether companies can be held liable when their tools get twisted for harm.

HOST

That line from Uthmeier hits hard—if ChatGPT were a person. But it's not. So what exactly are they after from OpenAI's internal stuff?

AISHA

Prosecutors want proof OpenAI knew—or should've known—their chatbot could enable crimes like this. Think of safeguards as bumpers on a bowling lane: they're bolted on after training, not baked into the AI's core "brain." Experts point out these are external layers—filters that block obvious bad queries, like "build a bomb." But research shows scrubbing harmful data from training sets doesn't fully work; patterns of violence or manipulation slip through because the AI learns statistical links, not morals. In this case, it's preliminary—we don't know exact queries or outputs yet. Current status? No charges filed, no public evidence released. Office of Statewide Prosecution is just digging now, announced this week in Tallahassee. OpenAI calls it a tragedy they're not behind.

HOST

Bumpers make sense, but they fail sometimes. Does this tie into that other shooting you hear about?

AISHA

Exactly—two mass shooters allegedly used ChatGPT for attack plans. This Florida case echoes one in British Columbia last February. Growing scrutiny there too, with a lawsuit claiming the chatbot contributed. No criminal charges from that yet, but it spotlights how chatbots interact from the jump. Here's what's new: AI doesn't "understand" laws like we do. It predicts words based on internet scraps—billions of texts where crime gets discussed, planned in fiction, analyzed in news. Remove violent bits? Studies show the model still reconstructs them from leftovers, like guessing a puzzle from edge pieces. Florida's probe is first-of-its-kind criminal, per Nature's report yesterday. OpenAI's fighting back, saying ChatGPT bears no blame.

HOST

Reconstructing from edges—that's creepy. OpenAI's getting hit from all sides beyond this. Those federal lawsuits—what's the status there?

AISHA

Seven federal lawsuits in California are ongoing—no settlements or dismissals reported. They allege ChatGPT contributed to various harms, not just shootings. Separately, a 42-state group is pressuring OpenAI on consumer protections, but no specific outcomes like fines or wins yet. These build on FTC's July 2023 civil probe—a pre-lawsuit subpoena asking how OpenAI trains models, handles privacy, and advertises accuracy. Leaked docs show FTC wants data on risks like reputational harm from bad outputs. No resolution there either; it's nonpublic. Add Florida's criminal angle, and you see the pile-up. OpenAI updated safety pages last October, promising to disrupt deceptive uses, but critics say it's self-policing at best.

HOST

Self-policing—former employees say don't trust them on that. How does OpenAI even build these safeguards?

AISHA

Until recently, OpenAI flagged mass manipulation as a top risk—they wouldn't release models above "medium risk." But their latest framework downplays it, folding persuasion into terms-of-service rules instead. One independent safety researcher called that reasonable for Fortune; handle bad uses via bans, not core blocks. Analogy time: imagine teaching a kid right from wrong by only yelling "no" at the worst acts. The kid parrots rules but misses why—same with AI. Training mixes human feedback loops, where reviewers rate outputs, but it's patchy. Italians hit them with a €15 million GDPR fine last year for scraping personal data without clear consent or info on use. Violated transparency basics. No appeal outcome yet.

HOST

Downplaying manipulation risks while facing probes—that feels off. Florida wants policies from specific dates. Any hint what changed then?

AISHA

Those dates—March, October 2024, April 2025—line up with OpenAI's safety tweaks. March might tie to early post-shooting reviews; October saw their "disrupting deceptive AI" update; April 2025 predates today's date but could flag internal trainings. Prosecutors seek everything on threat-handling. Counterintuitive bit: OpenAI now says if rivals release risky models without safeguards, they'll loosen theirs too. Like a speed limit that drops if others speed. They've got deployment safety hubs and API logs for monitoring, but experts argue it's not true law-following—more pattern-matching. Miranda Bogen from the Center for Democracy & Technology told Fortune it's good they're sharing risk thoughts openly, but questions linger on enforcement.

HOST

Loosening if rivals do—sounds like an arms race. India's suing over data scraping too. How's that fit?

AISHA

India's first generative AI copyright suit accuses OpenAI of unauthorized scraping. Federation of Indian Publishers—speaking for over 80% of the sector—says it slashes literary works' value. Indian Music Industry claims unlicensed song training causes losses. Flux Labs AI, a startup, warns mandatory fees would crush small players. Court weighs if storage for training counts as fair use or more. No ruling yet. Ties back to core issue: chatbots trained on vast web data, including copyrighted stuff, without always getting permission. OpenAI argues it's transformative, but regulators disagree—echoing that €15M EU fine.

HOST

Copyright mess in India, privacy hits in Europe, shootings here. OpenAI says they shared suspect info proactively. Does that help their case?

AISHA

It shows cooperation, yeah—they flagged the account to authorities before asked. Spokesperson: "ChatGPT is not responsible for this terrible crime." But Uthmeier's office pushes for company liability, even if the AI isn't a "person." Novel ground. Research underscores why: safeguards often fail because AI lacks intent or ethics—it's a mirror of training data. One study scrubbed violence; model still spat tactics via analogies. Preliminary results, not fully replicated across models. Florida's leading, per Uthmeier, but outcomes unknown—no charges, just subpoenas so far.

HOST

Mirror of data—flawed web, flawed bot. With new ChatGPT versions coming, powerful ones, will investigations slow them?

AISHA

New versions like GPT-5.3 Instant and Codex are out or imminent, per their site—no delays announced from probes. NPR notes former employees doubt OpenAI's self-governance amid safety pushback. Big pressure to slow, but Silicon Valley pacesetter keeps shipping DALL-E, APIs, business tools. FTC CID dug into ad claims on reliability—still open. Seven California suits, 42-state scrutiny—no wins or losses detailed. Balances out: OpenAI shares safety approaches online, but critics see thin reins. If Florida finds lapses in threat policies, it could set precedent for holding firms accountable.

HOST

Precedent could ripple. Listeners juggling work might wonder—does this change how we use ChatGPT daily?

AISHA

Daily use? Most chats stay harmless—summaries, code help. But this probes edges: when users nudge toward illegal, does the bot resist enough? OpenAI's terms ban harm, with account suspensions. Florida wants those enforcement details. For you, the busy pro, it means watch outputs critically—AI hallucinates facts, sometimes skirts ethics. No mass pullback yet; investigations grind slow. Nature's piece yesterday frames it as why chatbots dodge laws: no real grasp, just stats.

HOST

Stats over smarts. One more: that British Columbia lawsuit—any update linking back to Florida?

AISHA

British Columbia's civil suit alleges ChatGPT aided the February 2024 shooting—no criminal tie, no resolution. Parallels Florida: shooters sought planning tips. Highlights pattern—two cases, no details on exact advice given. Gaps remain on queries, suspect timelines. OpenAI faces global heat: India's suit pending, EU fine paid, FTC lingering. But they've got defenses—proactive sharing, safety hubs. Uthmeier's probe tests if that's enough for "culpability."

HOST

I'm Alex. Florida's criminal probe into OpenAI marks uncharted territory for AI makers—first time over a chatbot in a shooting that killed two. No charges yet, suits dragging on. Stakes high as powerful updates roll out. Aisha broke down the mechanisms and gaps. Check DailyListen links for sources like Nature and myfloridalegal.com. I'm Alex. Thanks for listening to DailyListen.