Curie is an AI assistant with a chemistry brief. This page says what that means in practice, including the parts that are inconvenient for us.
Final legal, tax, and operational launch verification is still pending. New Checkout is unavailable until the operator records approval. This policy was last updated 2026-08-03.
Curie is software, not a person, and nothing it writes has been reviewed by a chemist before you read it. It has a name and a warm tone because that makes it pleasant to use, not because there is anyone there. Substantive answers are generated fresh, in the moment, by an external AI service; the app can also show fixed local safety, refusal, or unavailable-output messages.
Using the assistant sends the text of your question to the configured external AI service. Aliquot does not use your prompts to train its own assistant. Privacy and Subprocessors identify the current product paths and the limits of what Aliquot can verify from source code.
If your institution cannot accept the active service’s current processing terms, do not use Curie. Every calculator, the converter, the safety data, and the reference tables remain available without an AI request.
It is sent to the configured external AI service, which returns the reply as a stream of text. If you selected Ask Curie on a calculator, the request can also contain that calculation’s inputs, results, and working. We do not write your question or the reply to a database, a log line, or an analytics event, and once the stream finishes our copy no longer exists. Your own browser keeps the conversation for the life of the tab so a reload does not lose it, and closing the tab discards it.
We keep counts only: that a question happened, roughly how many tokens it used, and which calculator topic the answer touched. Privacy has the table.
Because the text goes to a third party, do not paste anything confidential, personal, or patient-related into it. Once it leaves, that provider’s retention rules apply and not ours.
| What it does | What to do about it |
|---|---|
| States a wrong number with complete confidence. This is a characteristic AI failure, and the most dangerous one in a lab. | Take every number to the matching calculator, which is deterministic and tested. The assistant offers the link for exactly this reason. |
| Invents a plausible citation, product code, or supplier value. | Treat any specific reference it produces as unverified until you have seen the source yourself. |
| Misses an assumption: temperature, hydrate form, purity, salt form, valence. | Ask it what it assumed. State your own conditions explicitly in the question. |
| Answers a slightly different question than the one you asked. | Re-read your question in the thread. Ambiguity in, ambiguity out. |
| Stops mid-answer, or the connection drops. | The interface tells you when a reply was cut short rather than presenting it as complete. Retry. |
| Refuses something legitimate. | The safety limits are deliberately conservative and will sometimes catch ordinary hazardous-reagent work. Rephrase, or use the calculators and the safety pages, which have no filter. |
It declines to assist with explosives and energetic materials, chemical or biological weapons, poisons or gases intended to harm anyone including the person asking, and illicit drug manufacture, regardless of how the request is framed. Fictional, academic, and hypothetical framings are refused exactly as direct ones are. It weighs the whole conversation rather than the latest message, so a sequence of individually harmless questions that assembles a prohibited route is also declined.
If a message suggests someone may be considering self-harm, it does not provide lethal-dose or toxic-exposure information under any framing, and points to emergency help instead.
It also does not give medical, veterinary, clinical, or dosing advice. It will say so and suggest a qualified professional.
Explaining why a preparation behaved oddly, choosing between buffers, sanity-checking an approach, translating between conventions, and the ordinary bench questions that are too contextual for a calculator and too small to interrupt a colleague for. Used that way, with the numbers verified in the calculators, it is genuinely useful. That is the honest scope.
Every question costs real money to answer, so every plan carries a finite monthly allowance plus hourly and daily ceilings. They are enforced on the server. This is also why the assistant can be switched off at short notice if cost or abuse demands it: an unbounded AI budget on a small operation is not generosity, it is a service that eventually disappears for everyone. See Terms.
The assistant generates text. It does not score you, rank you, decide your price, or gate anything based on what you asked. There is no automated decision-making with legal effect within the meaning of Article 22 GDPR.
A specific report, the question and what the answer got wrong, is genuinely valuable and changes the product: bad answers in a recurring category get handled deterministically in a calculator instead. hello@aliquotchem.com.