Pinned Tweet
An AI answer with no source behind it is a guess.
Grounding fixes that: the model reads your documents before it answers.
Which raises the question most teams skip: which documents may it read, and where do they go?
Good plain-English guide in the reply.

ALT Diagram titled 'An answer without a source is a guess.' Left, From memory: dashed, empty source circles feed an answer; sources it remembers may not exist. Right, From your documents: a field of document dots, two ringed in red, behind a gate; only two pale dots pass the gate to the answer. Caption: the question becomes which files it may read. Botformed.
Your AI just gave you 12 cases for your brief. Names, holdings, links. It looks like research.
One question that's easy to skip: did it open them? 🧵
1
1
17
4/ Questions worth asking before relying on the list: which cases did it open in full? Which came from a summary? Can it show the passage and page?
If you represent yourself, the filing is yours. If you represent clients, you answer to them.
1
4
Sources:
Midpage, 6 Nov: midpage.ai/blog/opus-5-5-wit…
New York courts' AI guidance for attorneys: news10.com/wp-content/upload…
4
In Nature: AutoDNA, a lab where AI agents run DNA experiments on over 20 types of instruments. People restock supplies and service hardware.
In one test it matched expert-optimized enzymatic DNA synthesis yields.
For R&D teams, the question is what those agents may read.
30
A real case can still be the wrong answer.
One public tracker lists 2,266 court decisions where courts found or implied reliance on AI-generated material. 960 are tagged for real authority misrepresented, an error that can pass any "does this case exist?" check.
A thread 🧵
8/ Skipping the check costs. Flycatcher v. Affable Ave. (S.D.N.Y.): default judgment against the lawyer's own client, then $38,877.59 in fees and costs. Kleyman v. Kaloidis (N.Y. Sup. Ct.): a $10,000 sanction plus $36,511.35 in fees.
Experts too, as @nikiblack wrote.
1
9
9/ What grounding means in law:
1. Answer from the record, not memory
2. Run all three checks
3. Record who checked what, and when
4. Fail closed (Sachar's term): a failed check holds the document until someone resolves it
Full article, every source:
linkedin.com/pulse/real-case…
8
Botformed retweeted
AI agents enter the wet lab.
In new @Nature N&V article, I discuss recent advances in agents running automated labs and ideas on how to make these systems more reliable. nature.com/articles/d41586-0…
Grounding, RAG and tool use, explained by @cowshedstartup: cow-shed.com/blog/grounding-…
10







