Transparency
Our Positions on AI
Where we stand — plain language, no marketing speak.
Using LLMs to build products means taking a position on a set of real questions — about how the models were made, what they're good for, who benefits, and what happens to your data. Most companies answer those questions with marketing copy. We'd rather say what we actually think.
These are CF's honest stances. When something changes — a product ships that crosses a line we said we wouldn't cross, a position gets refined by experience — this page updates first.
Training data provenance
The models we use were trained on data scraped at scale without meaningful consent from the people who created it. We think that was wrong. If we had the resources to build our own models from the ground up, we'd do it differently — consented datasets, fair compensation, the whole thing.
We don't have those resources. And the models exist regardless of what we do next.
Our position is this: refusing to use them doesn't undo the harm — it just determines who benefits from the capability that already exists. We didn't build these tools and weren't asked whether they should exist. But we get to choose what they're pointed at. We point them at people navigating systems that were already failing them. That's not an absolution. It's a choice we're making with open eyes.
LLM reliability and appropriate use
LLMs are useful for some things and dangerous for others. We use them for drafting, summarising, and generating starting points — tasks where a wrong answer costs a revision, not a disaster.
We don't use them to make decisions, assess eligibility, or produce outputs that go directly into high-stakes systems without human review. Every LLM output in CircuitForge products is a draft that a human approves before it goes anywhere.
Where deterministic processes exist, we use them. LLMs fill the gaps — they don't own the critical path.
Labor and what we automate
Most of what CircuitForge automates is overhead — the bureaucratic friction of navigating systems that were deliberately made opaque and exhausting. Writing your fifteenth cover letter isn't a livelihood; it's a tax on job seekers. Sitting on hold for two hours isn't a skill; it's attrition. We're automating the friction, not the work.
Our honest read of what these tools actually are: a means to help humans accomplish more, do more, learn faster. They represent a meaningful step forward in how useful machines are at interfacing with humans — translating intent into structured output, drafting from rough notes, processing hours of reading into a few seconds, checking for errors in code and writing. That's a real shift, but it is not judgment, creativity, or the kind of labor that actually matters. Like any tool, they can build or break. A hammer builds a house or hurts a person. The house shelters people or becomes a prison. Intent and method of use determine what you get with any tool. We're choosing to use this one to remove obstacles — the grinding overhead that was already costing people time they don't have.
That said, we won't pretend every product we build will stay cleanly on that side of the line. As CF grows, some tools may touch workflows that involve real human labor. When that happens, we'll say so plainly and think carefully about the design. If it does, our aim is to eliminate the parts people don't want to do — not to replace the people doing them. Our intention is to help people, not replace them.
Commercialisation
CircuitForge is a business because it has to be. The BSL (Business Source License) on AI features and the paid tier for cloud inference exist because the alternative is the project dying. We're not chasing growth, scale, or an exit. The goal is to cover costs, sustain a team, and keep building.
What we're explicitly not doing: taking funding that creates growth obligations that corrupt the mission. The difference between a crowdfunding backer and a venture fund is this — a backer wants the product to exist; a fund needs a 10× return in five years. Those are not the same incentive, and they don't produce the same product.
The free tier is real and fully functional. The open-source layer is real and MIT-licensed. The only thing behind a paywall is use of our own hardware and managed services. We'd prefer a world where this didn't require a business at all. It does.
Scale and growth
The difference between CF and most AI companies isn't that we don't want to succeed — it's that we won't use the growth tactics most AI companies depend on: harvesting user data, engineering engagement, locking people into platforms. Those tactics work, but they work by treating users as a resource to be extracted. Our users are already being extracted from by the systems they're navigating. We're not adding to that.
Our growth model is simpler and more challenging: build tools people trust enough to recommend.
Energy and compute
The ecological cost of AI is almost entirely in pre-training foundation models — weeks of massive compute consumption at scales only a handful of organisations can afford. CircuitForge does train models, but not at that scale: we fine-tune smaller, task-specific models for targeted jobs. That's a fundamentally different energy profile. Running inference for individual use is genuinely small: a research session on local hardware draws about 75W for 5–10 minutes — less than charging a phone.
The real problem is scale and centralisation: millions of queries routed through hyperscaler data centres running on non-renewable grids. Our own cloud infrastructure runs on solar. The case for local inference isn't primarily ecological — it's about keeping compute distributed, keeping data off centralised servers, and not feeding the concentration of compute and energy into a handful of facilities. Those are privacy and resilience arguments as much as environmental ones.
We encourage local inference because it's better across multiple dimensions — not to assign individual users guilt for a few watt-hours.
Cloud privacy practices
When you use CircuitForge's cloud tier, your data is processed with the same privacy intent as local inference — just running on our hardware instead of yours.
We don't log personally identifiable information where we can avoid it. Where some data handling is technically unavoidable, we obfuscate it. We retain nothing longer than we need to complete the task. We are not building a behavioural profile of our users. We are not cataloguing your activity to train future models or sell to advertisers.
Any data we do collect requires plain-language informed consent — not a 40-page terms of service, not a pre-checked box. Informed consent means you can actually understand what you're agreeing to. If you need a lawyer to parse it, it isn't informed consent — it's legal cover. We write ours so a person can read it and know what they're signing up for.
If you ask us to delete your data, we will delete every iota of it. No friction, no retention period, no "we'll get to it." You ask, it's gone. This isn't a policy promise — we've built the endpoint. Deletion is a first-class feature of our platform, not an afterthought added to satisfy a regulation.
This isn't a privacy policy clause that lives in a document nobody reads — it's the design principle that shapes how we build every workflow. See our full Privacy Policy for the specifics.
Questions about how any of this applies to a specific product?