The Kitchen Sink for April 3, 2026: Legal Tech Trends

AI Hallucination Cases Just Topped 1,228 — And Courts Are Starting to Restrict Where Evidence Can Go

Doug Austin's weekly "Kitchen Sink" roundup on eDiscovery Today, published April 3, 2026, strings together several items that matter more read as a set than individually: the running tally of AI hallucination lawsuits reached 1,228, with 472 of those involving lawyers directly; a Google Analytics sunset produced a spoliation sanctions case after an IT employee preserved data without direction from counsel; a new study found sycophantic AI tools make people more confident they're right and less willing to resolve disagreements; a court addressed a pro se litigant's AI use as work product alongside a separate protective order barring uploads of confidential material to LLMs; and another piece pegged current AI infrastructure spending at $650 billion.

Why does the hallucination case count matter to e-discovery buyers?

A running tally like this is a leading indicator that courts are treating AI output errors as sanctionable conduct, not just embarrassing footnotes, across more litigation each month.

472 of the 1,228 cases involve lawyers themselves relying on fabricated citations or filings, not just AI used by opposing parties. That means the exposure isn't hypothetical for firms adopting GenAI research or drafting tools; it's already generating sanctions orders. Austin's post cites commentator Jerry Lawson's framing of TAR as "the old, reliable workhorse" against GenAI's newer, unproven track record — memorable, but the real point underneath it is that TAR earned judicial trust over roughly a decade through documented, measurable accuracy. GenAI tools haven't put in that time yet.

What does the protective-order ruling mean for where evidence can go?

A court approved barring uploads of confidential material to LLMs, while a separate ruling held that a pro se litigant's AI use counted as work product.

Together those are two signals that courts are drawing sharper lines around what's allowed to touch a third-party model, and that crossing those lines carries privilege consequences. Michael Berman, discussing the decision on the EDRM blog and cited in the roundup, flagged the holding that routing electronic interaction through third-party systems doesn't automatically forfeit a reasonable expectation of privacy — a narrow, specific finding that will almost certainly get tested and narrowed in the next case, not a blanket rule to rely on.

What should you actually ask your own AI-in-discovery vendor?

Ask whether privileged evidence ever leaves your environment, whether it trains any shared model, and who can prove it didn't.

Nothing in this roundup measures any specific vendor against those questions — it's case law and commentary, not a benchmark study. That's the gap: if a platform can't show single-tenant isolation, your own cloud keys, and a documented chain of custody, "we don't train on your data" is a policy statement, not a technical guarantee. Given where courts are heading on LLM uploads, that distinction is now a diligence item, not a nice-to-have.

Does the $650 billion AI infrastructure figure affect a legal team's budget?

Only indirectly — it describes enterprise-wide capital spending trends, not e-discovery-specific costs or return on investment.

The figure comes from a broader piece Austin links to approvingly, comparing the current AI buildout to the early days of cloud computing. That's a reasonable macro signal for multi-year technology planning, but it says nothing about what any particular AI discovery tool costs to run, host, or defend in front of a judge — questions buyers still have to settle vendor by vendor.

Frequently asked questions

Does "472 lawyer-involved hallucination cases" mean 472 attorneys were sanctioned?

No. The count, as reported, covers cases where lawyers were involved with AI hallucinations in some capacity; not every instance produced sanctions.

Did the roundup name which e-discovery platforms are affected by the protective-order ruling?

No. The cases discussed concern LLM upload restrictions and work-product findings generally, not any named e-discovery vendor or product.

Is TAR being replaced by GenAI in this coverage?

No. The commentary treats them as complementary, using TAR's decade-long accuracy record as the bar GenAI tools still have to clear.

Source: eDiscovery Today — "The Kitchen Sink for April 3, 2026: Legal Tech Trends," by Doug Austin.