Laboratory
Artificial intelligence, without faith
I use these tools every day and I have built systems on top of them. That obliges me to be exact about two things: what they actually deliver, and where they cannot be used. Two topics — Claude, by Anthropic, and all the rest.
Claude, by Anthropic
Claude is a language model developed by Anthropic. The common description — "it chats and writes text" — describes the toy, not the instrument. What changes the nature of the work is what came after the text: the ability to operate on my own files, execute commands, query external systems and sustain a long task through to the end.
Context
How much material the model can hold in front of it at once. It is what makes it possible to put an entire case file under analysis rather than scattered extracts — and it is also the practical ceiling that forces a slice-by-slice reading strategy.
Tools
The model does not merely answer: it executes. Reading a file from disk, running a script, querying an API, writing a document. The answer stops being an opinion about the data and becomes an operation on the data.
MCP
Model Context Protocol — an open standard published by Anthropic for connecting a model to external systems (databases, repositories, services) without bespoke integration for each one. It is what makes it viable to wire AI into infrastructure a firm already has.
Agents
Instead of one question and one answer, an objective and a loop: plan, execute, verify, correct. It is how a task of several hours — mapping a caseload, reconciling deadlines — becomes delegable.
Skills
Standing instructions that encode a method: how this firm drafts a pleading, what verification is mandatory before citing case law, how to read a court PDF at minimum cost. Tacit knowledge becomes an executable procedure.
Verification
None of this dispenses with checking. The real gain is not trusting more — it is being able to check faster, because the material arrives organised and with its source reference preserved.
What I built with it
Litigation management integrated with DataJud
A platform that tracks docket movements across a caseload of roughly one to two thousand matters through the National Council of Justice’s public API, extracts deadlines, sends notifications and maintains a file for each case. Web and mobile.
Reading court records at minimum cost
A procedure that turns a court PDF of hundreds of megabytes into a map of filings drawn from its bookmarks, extracts only the slice required, and applies OCR solely to pages that are images — preserving the page anchor needed for citation.
Assisted drafting of pleadings
A method that imposes the firm’s structure, mandatory verification of case law against the official source, and the form the courts require — including the opening summary demanded by article 343-A of the Superior Court of Justice’s Internal Rules.
Personal-data anonymisation
Removal of identifiers before any assisted analysis, so that processing personal data is not a side effect of using the tool.
The other AIs
There is no single tool. There is a fit between the task, the risk and where the data ends up. I use more than one, and the choice shifts with the matter — anyone selling a single answer is selling, not analysing.
Family I
Conversation and reasoning
The general-purpose models. This is where the commercial fight is concentrated and where the differences, in practice, are smaller than the marketing suggests. The choice usually turns on integration, cost and data policy — not raw capability.
- GPT — OpenAI The most widespread, with the largest integration ecosystem
- Gemini — Google Integrated with the Google environment and strong on very long context
- Grok — xAI Conversational, coupled to the X social network
- DeepSeek Open weights and low cost, from a Chinese laboratory
In litigation A good share of what reaches a case file today has passed through one of these. When examining a document, the question is no longer whether AI was used, but what part it played in producing that particular filing.
Family II
Search and grounded synthesis
Tools that do not claim to know, but to find and summarise with a declared source. For legal work they are more useful than pure chat, because they hand back the path to the source.
- Perplexity Web search with a cited answer and a link to the origin
- NotebookLM — Google Summarises and answers only over documents you upload yourself
- Google Labs An umbrella for experiments — useful, and unstable by definition
In litigation The citation shown still has to be opened. "It has a source" is not the same as "the source says that" — and I have checked cases where the link existed and did not support the assertion.
Family III
Agents — execution, not conversation
Rather than answering, they act: they take an objective, plan, operate on files and systems, verify and correct. This is the category that changes office work, and also the one that demands the most governance.
- Manus Autonomous agent for multi-step tasks
- Terminal agents Operate directly on files, repositories and local systems
In litigation An agent that acts needs explicit limits: what it may touch, what requires confirmation, what is logged. Without that it is not automation — it is operational risk with a pleasant name.
Family IV
Synthetic media — the family that produces the evidence I examine
Generation of video, voice and avatars from text. Forensically this is the decisive family: it is the source of the deepfake filed in the record, the audio attributed to someone who never spoke, the video of a scene that never occurred.
- Veo — Google DeepMind Video generation from a text description
- Hailuo — MiniMax Video generation, widely adopted on social platforms
- HeyGen Video avatars, voice cloning and lip-synced dubbing
- Image generators Photorealistic creation and editing, including of faces and documents
In litigation Quality has already passed the untrained eye — and, in many cases, the trained one. The examination therefore moves from appearance to trace: metadata, provenance, generation artefacts and the route by which the content reached the record.
Family V
Local execution and open models
Models running on your own machine, with no connection. Less capable than the frontier services, with one property none of them offers: the data never leaves the equipment.
- Open-weight models Run on your own server or workstation, sending nothing out
- DeepSeek and the like Also available for local execution, alongside the cloud service
In litigation For material under seal, expert evidence and confidential documents, this is usually the only defensible option — and it is the technical answer to the question "can I use AI without breaching data-protection law?".
Family VI
Specialised legal platforms
Products aimed at litigation: importing case records, extracting the questions put to the expert, drafting, case law. They deliver a ready-made workflow and save configuration.
- AI-enabled practice suites A closed workflow, from the case file to the draft
- Expert-examination platforms Import of the record and draft answers to the questions put
In litigation The hidden cost is the criterion: it sits inside someone else’s black box. And it is precisely the criterion the expert must be able to explain when the report is challenged.
This survey ages quickly. Names change, products are discontinued and capabilities jump from one month to the next — what is described here is the function each family performs, because function outlasts brand. If the decision matters to a case, confirm the current state before relying on it.
What they all have in common — and nobody advertises
They all fabricate with conviction. A language model produces the most probable text, not the true text — and a non-existent judgment, complete with docket number, rapporteur and perfectly formatted headnote, is highly probable. Every citation has to be opened at the official source. Without exception.
They all process data. Sending a pleading, a report or a personal document to a cloud service is processing personal data, with every consequence that follows under data-protection law. Either it is anonymised first, or it runs locally, or it is not done.
None of them answers for anything. There is no appointment, no oath, no criminal liability for a mistaken finding. The legal system requires an identifiable responsible person, and that person is always a human being.
What I delegate, and what I do not
Verdict
Yes — this is where the machine beats the human
Why
Volume and search are the strong suit. The expert checks the page reference before citing; models get numbering wrong easily.
When the AI becomes the object of the examination
The question arrives with growing frequency: was this document, this audio, this image produced by artificial intelligence? The honest answer begins by discarding the method everyone tries first.
Stylometry does not solve it. Detectors that classify text by how much it "looks like AI" produce false positives and false negatives at a rate far too high to support an expert conclusion — and they penalise precisely those who write in an organised way. A report resting on that does not survive a technical challenge.
What supports an examination is the material trace: file metadata, edit history, compression and generation artefacts, capture inconsistencies, provenance declared in the file itself, and the route by which that content reached the record. It is chain-of-custody work, not an opinion about style.
- 01 Metadata and provenance declared in the file
- 02 Edit history and coherence of the dates
- 03 Generation artefacts in image, voice and video
- 04 Consistency with the capture device claimed
- 05 Route into the record and chain of custody
- 06 Comparison with the original, where an original exists
The thesis
The right question is not whether AI will replace the expert. It is what, specifically, a system cannot do in an examination — and what that means for whoever signs the report.
When a court rejects a report produced by generative AI for want of minimum epistemic reliability, it is not merely discarding a document. It is reaffirming that the human credential — trained, accountable, auditable — is what validates a technical conclusion within the proceedings. The technology lets me work faster and look at more material. It does not sign.
Deployment, method, lectures
If your organisation is going to use this, someone has to know where the limits are.
Workflow design, usage policy, data-protection compliance, choice of tool by data type and team training — for law firms, companies and institutions.