Brian Reed: Hey — glad you're here, because I genuinely could not figure out how to frame this one.
Eliza Ward: The SEC story?
Brian Reed: Yeah. Because there are actually two things happening and they pull in opposite directions. The SEC stood up an Office of Artificial Intelligence and Analytics — their own AI surveillance unit, built to catch manipulation.
Eliza Ward: And simultaneously they're prosecuting people for faking AI.
Brian Reed: Which is a totally different problem. The Nate, Inc. case — Albert Saniger allegedly told investors his shopping app ran autonomously, no humans touching it. That was the core claim. He raised over forty million dollars on it.
Eliza Ward: April 2025. And SDNY filed a criminal indictment alongside the SEC action — that detail matters because it signals how seriously they're treating the fraud.
Brian Reed: Right — so the enforcement arm is moving. But I keep asking: is the AI surveillance office catching the people who are actually using AI to manipulate markets? Or is it just — wait, no — is it mostly catching the people who lied about having it?
Eliza Ward: Those are very different detection problems. And I'm not sure the office is equipped for both at once.
Brian Reed: And that gap — that's actually what the Kubient case exposed. September 2024, Paul Roberts gets hit with SEC enforcement and DOJ criminal charges simultaneously. First time a public company executive gets tagged this way on AI washing specifically. And the SEC didn't find him by running surveillance. Someone already knew.
Eliza Ward: The SEC had to name the pattern before they could prosecute it. The Office of AI and Analytics didn't detect Kubient — it categorized what already arrived.
Brian Reed: So the office is a label, not a dragnet.
Eliza Ward: Maybe. I mean — wait, that's not quite fair either. GenesisAI, August 26, 2026, settled action, Archil Cheishvili facing fifty thousand dollars disgorgement plus a permanent injunction for misstating AI marketplace viability. That's a small number. That's not deterrence at scale, that's a footnote.
Brian Reed: Fifty thousand dollars. Against someone who was marketing an AI marketplace.
Eliza Ward: Right. And Cornerstone Research and the Stanford Securities Class Action Clearinghouse are reporting AI-related class actions more than doubled from 2023 to 2024. So private litigation is accelerating way faster than enforcement is.
Brian Reed: So the question isn't whether the doctrine exists — it clearly does, Kubient, GenesisAI, Nate Inc., those are real cases with real names. The question is, let me see — are these the two percent who got unlucky, or are they actually representative of what's being caught?
Eliza Ward: And the SEC can't answer that. Because if they could quantify the gray zone — the overclaiming that never crosses into Saniger territory — they'd already be prosecuting it.
Brian Reed: So what's the concrete thing to watch here — is it whether the GenesisAI final judgment actually gets entered, or whether any of these cases produce a conviction rather than a settlement?
Eliza Ward: Whether it gets entered — yeah, that's real. But that's actually the smaller question. The real one is whether the whole framing of the SEC's AI office is just wrong from the start.
Brian Reed: Which framing specifically?
Eliza Ward: That it closes a gap. The take circulating right now — that the Office of AI and Analytics gives the SEC a structural edge — I don't think that holds. Here's the actual problem: a bank's fraud team and a card skimmer can buy the exact same detection software. Having the tool doesn't mean you see the crime before it clears.
Brian Reed: Okay, so both sides have better shovels. But doesn't the SEC have more data?
Eliza Ward: Same data. That's the whole point. Regulators and bad actors are drawing on the same underlying market data, the same AI capabilities. The advantage goes to whoever acts first. And in modern markets, a transaction can propagate — it clears, it settles — before any detection system flags it. That's the Speed-Fraud Paradox. AI narrows that window. It does not close it.
Brian Reed: And Gensler basically admitted that framing, didn't he — I mean, at the MIT AI Policy Forum he said AI was 'every bit as transformational as the internet.' That's not a regulator saying we've got this. That's someone saying — wait, actually — we're already behind a force reshaping the market.
Eliza Ward: Right. And he's gone — departed early 2025. We can't attribute that framework to whoever's running the SEC now. So we have an office built on a catch-up logic, articulated by a chair who's no longer there.
Brian Reed: And the cases we can actually point to — Nate, Kubient — those weren't caught algorithmically. Someone investigated. The AI office categorized them after.
Eliza Ward: Which is — look, that's triage, not adjudication. And there's a reason that distinction matters legally that we haven't gotten to yet — the part about what happens when the SEC has to explain in court how their model reached an output. It gets genuinely harder there.
Brian Reed: And that's — yeah, that's the thing that actually keeps this from being a staffing problem. Because even if the SEC had ten times the investigators, the AI output still can't walk into a courtroom and explain itself.
Eliza Ward: Right. AI as triage, not adjudication. The model flags something — prioritizes it for human review — but it cannot prove intent. And intent is what securities fraud requires.
Brian Reed: So the flag is real but the case isn't — not automatically.
Eliza Ward: And it cuts the other direction too — that's the part I think gets missed. A defendant whose strategy runs inside an algorithmic black box can say: the outputs weren't foreseeable. I didn't direct the manipulation, the model did. The same opacity that constrains the SEC is — wait, actually — it's a shield for the defense.
Brian Reed: Hold on. That's a viable legal argument?
Eliza Ward: It's untested at scale. But if the SEC's own model can't produce a human-legible explanation of why it flagged a trade, how does a prosecutor establish that a defendant's algorithm — just as opaque — was acting intentionally? Both black boxes, neither one opens.
Brian Reed: So make it concrete — like, what does this actually look like? Not the trader. The person one step back.
Eliza Ward: Thursday afternoon, compliance officer at a mid-size fund. Her firm's ML model flags a microcap options pattern — looks like pre-announcement positioning. She files the report internally, it moves into the SEC's Office of AI and Analytics as an aggregated signal. Three weeks later. The position has already settled. Everyone did exactly their job. The manipulation cleared.
Brian Reed: And that's not anyone failing — that's the evidentiary architecture itself. The latency isn't a bug in the process, it's baked into how the signal travels. What's the concrete thing to watch for whether this actually gets fixed — or at least tested?
Eliza Ward: Whether it scales. That's the signal. Not the GenesisAI judgment, not whether Saniger pleads — the thing to watch is whether any future case targets AI-enabled manipulation. Not someone lying about having AI. Someone actually using it to move a market. Those require completely different evidentiary tools, and so far we don't have a single enforcement action that gets there.
Brian Reed: And I mean, I feel it too, but it feels almost too clean to say it. Like, Kubient, Nate, GenesisAI — those are real cases, real names, real penalties. But they're all about lying about AI. None of them are about AI doing the bad thing. So the doctrine that exists and the problem that's actually coming... they're not lined up yet.
Eliza Ward: Not even close to lined up.
Brian Reed: So where does that leave market integrity as a concept — I mean, actually, I don't think that's answerable right now. What I can say is: if the speed of a transaction permanently outpaces the speed of a legal case, the SEC's AI office being reactive isn't a staffing failure. It might just be the architecture. And I don't know what tool — AI or otherwise — closes that.
Eliza Ward: Neither do I. And I think that's the honest place to stop.