Which models are used
Lensable does not train or run its own models. It routes your selection and question to a third-party provider that you choose:
- Lensable Auto — reads the selection and picks a provider for you
- ChatGPT (OpenAI)
- Claude (Anthropic)
- Gemini (Google)
- Grok (xAI)
- Perplexity
Some actions do not offer that choice, because they are not answering a question you asked about a selection: the Saved Highlights summary, and — where following is available on your account — the three steps described under Following below. Those use Gemini.
How Auto decides
Auto looks at what is in the selection before it routes. A chart with axis labels goes to a model that reasons well about numbers. A photograph goes to one that sees well. A citation or a claim that needs checking goes to one that searches.
Auto is a convenience, not a black box you are stuck with. You can override it for a single question or set a provider permanently, and the panel shows which model answered.
Following, where it is available
Where following is switched on for your account, three steps use Gemini. None of them is answering a question you asked about a selection, which is why the provider selector above does not apply to them.
- Suggesting a topic name. Runs only when you ask for a suggestion while adding a topic. It receives the text you typed and returns a proposed name and a one-line description of the subject area. Nothing is stored by the request and you can ignore what it returns.
- Judging which retrieved items are relevant. A keyword rule settles most items with no model involved. Only the ones it cannot settle go to Gemini, in one batched request per topic per retrieval run, carrying the topic name and each undecided item's title and excerpt.
- Writing a brief. The items retrieved for a period are sent to Gemini to write a short brief that cites them. Claims it cannot tie back to a retrieved source are dropped before the brief is stored.
All three read either publicly retrieved third-party material or text you typed yourself. None of them reads your Saved Highlights, saved items or search history, and none is used to work out an interest you did not state. Briefs are written from third-party sources and carry the same caveats as any other model output — read the cited source before relying on one.
What AI answers cannot do
Answers are generated by statistical models. They are frequently useful and sometimes confidently wrong. In particular:
- An answer may be fluent and still factually incorrect.
- Product identification is probabilistic. Image quality, cropping and catalogue coverage all affect it.
- A model may describe something that is not in your selection, or miss something that is.
- Models have knowledge cutoffs and may not know about recent events.
Treat Lensable's output as a lead to verify, not as a source. Do not rely on it for medical, legal, financial or safety-critical decisions. Confirm any product, price or availability on the seller's own site before you buy.
Labelling
Anything generated by a model is shown inside the Lensable panel and attributed to the provider that produced it. We do not present AI output as human-written editorial, and we do not blend it into page content so that it looks like part of the site you are reading.
Training
We do not use your selections or questions to train models. Providers apply their own policies to data they receive through their APIs; where a provider offers an API tier that excludes training by default, we use it.