When AI Screws Up, Nobody Knows Who to Blame — and That’s Becoming a Real Problem

AI is getting more capable by the week. It’s booking your travel, drafting your contracts, and soon it may be making decisions on your behalf without you lifting a finger. But three stories today point to the same uncomfortable truth: the world around AI isn’t keeping up.

When a self-driving car crashes, we at least have a rough framework for who might be responsible. When an AI agent — a system that can take actions autonomously, like sending emails, making purchases, or executing code — causes harm, things get murky fast.

MIT Technology Review reported today that legal experts and technologists are increasingly alarmed by a gap at the center of AI law: nobody clearly owns the consequences when these systems go rogue. Is it the company that built the AI? The business that deployed it? The person who gave it instructions? Right now, courts and regulators don’t have a settled answer.

Think of it like this. If a contractor builds a faulty staircase and someone falls, we know how to assign blame. But AI agents are more like contractors who designed themselves, were hired by someone else, and then made independent decisions on the job. Each layer of that chain has a plausible argument for pointing at someone else.

For everyday people, this gap is already relevant. If an AI system managing your insurance claim makes a costly error, or an automated hiring tool unfairly rejects your application, your legal options are genuinely unclear today. That’s not a hypothetical — those systems exist now.

Why this matters: Accountability is what gives legal systems their teeth. Without it, there’s little pressure on anyone to make these systems safer.

“It’s unclear who should be held responsible when AI systems cause damage or break the law.”

A Startup Just Raised $25 Million to Catch Fake Voices

Voice deepfakes — recordings where AI has been used to convincingly mimic a real person’s voice — are no longer a sci-fi concern. Scammers have already used them to impersonate executives and authorize wire transfers. Fraudsters clone relatives’ voices to trick elderly people out of their savings.

According to TechCrunch, Modulate has raised $25 million to fight back. The company builds tools that analyze audio recordings to flag manipulation, detect synthetic speech, and identify fraud in real time. Their technology is aimed at businesses that rely heavily on phone-based communication — think banks, insurers, and call centers.

The core idea is essentially a lie detector for audio. Just as your email provider scans for spam signals you’d never notice, Modulate’s tools look for acoustic fingerprints left behind when AI generates or modifies a voice. Real human speech has subtle irregularities; synthetic speech often doesn’t.

For you, this could eventually show up as a layer of protection on your bank’s fraud detection system, or as a tool your company uses before acting on any voice instruction. The stakes are high enough that $25 million in funding looks like a reasonable bet.

Why this matters: Voice is one of the last things we instinctively trust. Once that trust erodes at scale, the damage to how we communicate — in business and in families — is significant.

“Modulate raised $25 million for voice deepfake and fraud detection.”

TechCrunch Disrupt 2026 Is Almost Here — Last Call for Exhibitors

TechCrunch Disrupt 2026 runs October 13–15 in San Francisco, and according to TechCrunch, the deadline to reserve an exhibit table falls on Friday, October 2 at 11:59 p.m. PT. For early-stage startups looking to get in front of investors and potential partners, this conference is one of the year’s most concentrated opportunities to do exactly that.

The event draws thousands of attendees from across tech, with a particular focus on AI companies this cycle. An exhibit table isn’t just a booth — for a small startup, it can be the room where a key partnership begins. That said, it comes with a real cost, and not every stage of company benefits equally from a crowded conference floor.

For founders weighing the decision, the question isn’t just whether the conference is good. It’s whether their product is ready to make an impression with people who will compare it to dozens of competitors in the same afternoon.

Why this matters: The companies that show at Disrupt often signal what investors are paying attention to this season — worth watching even if you’re not attending.

“Deadline to reserve a booth is Friday, October 2 at 11:59 p.m. PT.”

Also Happening in AI

Several smaller stories round out a busy Monday. LiteLLM, a tool developers use to connect applications to multiple AI models at once, released version 1.103.0 with added cryptographic verification for its Docker images — a quiet but meaningful security improvement. Over at Wired, a piece argues that AI agents are about to flood the workplace and most organizations have no real plan for integrating them. Insuretech startup Outmarket raised $34.5 million to build AI tools for the insurance industry, per TechCrunch — part of a broader wave of AI investment in traditionally slow-moving sectors. Atlassian CEO Mike Cannon-Brookes pushed back on doomsday predictions for SaaS businesses in a conversation with The Verge, arguing the so-called “SaaSpocalypse” hasn’t materialized. And on a more playful note, The Verge covered Engram, a new musical instrument from Thoughtful Things that turns AI hallucinations — the confident but wrong outputs AI sometimes generates — into distorted, experimental sound.

What to Watch

The liability question isn’t going away. Watch for the first major court case where an AI agent causes measurable harm to a private individual — that ruling will set a precedent that shapes how companies deploy autonomous systems for years. On the voice fraud side, keep an eye on whether detection tools like Modulate’s get folded into regulatory requirements for financial institutions, the way fraud screening already is. If they do, voice verification could become as routine as a PIN.