What we can investigate
Intelligence for a bespoke investigation.
The answers
What is AI saying about us?
We start from your brief: what you are working towards, and who will be asking. From it we write the questions, the way a bank, a journalist or a counterparty would put them. Each goes to seven AI systems, and every answer is kept word for word.
- You get
- Each finding, with how many of the seven AI systems raise it and in how many answers. A second company's AI checks every reading, and a person settles each disagreement.
The sources
What is it relying on?
AI works like a blacklight: it shows up sources and connections you would not otherwise see. A person goes through what the answers cite for patterns, obscure sources and links nobody had noticed. One page can be taken apart on its own: its history, who owns it and what it is linked to.
- You get
- The pages behind each finding, and the ones that deserve a closer look.
- What it does not show
- A page an AI cites is not proof that the page shaped its answer.
Web search on and off
Does it come from the live web?
We ask the same questions twice: once with the AI's web search switched on, once with it off. Then we count which findings appear far more often with it on.
- You get
- The findings that depend on what is online today, set apart from the ones that do not.
- What it does not show
- An observed difference on the days of the test. It does not prove where an answer comes from, or that changing a page would change it.
How it ended
Does it tell the whole story?
You give us the facts: the case, the appeal, the settlement. We count which AI systems mention each one, and whether they say how it ended.
- You get
- For each matter: which AI systems raise it, and which leave out the ending. Checked by hand before you see it.
Message penetration
Is our message getting through?
You give us your statement or your key messages. We count which answers carry each one, AI by AI, with an example in the AI's own words.
- You get
- For each message: where it has landed and where it has not. An answer that denies the message does not count.
Personal data tracing
What personal details does it give out?
For a private person we go through every answer for dates of birth, addresses, phone numbers, emails, and identity and financial numbers, and trace each one to the page the AI cited beside it.
- You get
- What is exposed, by which AI system, and where it points. The details themselves are withheld from every copy we send.
Languages side by side
Is it different in another language?
The same questions in two or more languages, with the answers and their sources set beside each other. Run so far in English, French, Russian and Arabic.
- You get
- What each language says that the other does not, and which sources each one leans on.
- What it does not show
- Translated questions have not been shown to be equivalent, and we say so beside the figures.
Re-check
Has anything changed?
The same questions, asked again later, so you can see whether the answers have moved.
- You get
- What is newly raised, what is no longer raised, and which sources have changed since last time.
A breaking story
A story is breaking. What is AI saying now?
Seven checks over three days. Which AI systems have picked the story up, which pages they cite, and whether your statement is reaching them. Your own sites can be set beside a competitor's.
- You get
- A message from your analyst when something changes, and a final briefing.
The evidence
Can we rely on the record?
Every answer is signed and independently timestamped shortly after capture. Anyone can check the archive offline, without us.
- You get
- The original answers and their cited pages, in an archive that is yours to keep.
- What it does not show
- It shows that the record has not been altered. It does not make what an AI said true.