Building AI Responsibly: An AI Company’s Honest Stance
We build AI for a living, so it would be easy to only sell the upside. Here is the honest version — what the real risks actually are in 2026, and how a builder chooses to act on them.
2026 has been the year the ground moved twice at once. Scientists published the first near-complete wiring map of a whole brain — the male fruit-fly central nervous system, roughly 166,000 neurons and ~125 million connections, mapped by Google Research, HHMI Janelia and the University of Cambridge (Cell, 3 September 2026), building on the 139,255-neuron female FlyWire connectome in Nature two years earlier. In the same window, AI systems began generating entire playable worlds from a prompt — from Oasis, a neural-network Minecraft with no game engine, to DeepMind’s Genie 3. We can now read biological intelligence and write synthetic worlds in the same news cycle — while understanding neither fully.
We build AI for a living. That gives us every incentive to only tell you the exciting half. This post is the other half: an honest account of what the risks really are, and how a company that ships AI chooses to act on them.
Start by taking the warnings seriously
The strongest case for caution doesn’t come from critics on the outside — it comes from the people building the frontier. In 2023, hundreds of researchers and executives, including OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and Turing Award winners Geoffrey Hinton and Yoshua Bengio, signed a one-sentence statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Amodei has written that he “think[s] and talk[s] a lot about the risks of powerful AI” even in an essay about its upside. More recently, in an opinion column for The Guardian (published ~14 September 2026), former Google DeepMind researcher Alex Turner argued the field is running “an extremely dangerous race towards superintelligent AI” — putting his personal estimate of an eventual AI takeover at roughly one-in-three. That figure is his own guess, not a measurement, and we treat it as such. But the direction of concern is broadly shared: even Demis Hassabis has called for research into AI’s threats “to be done urgently” and for “smart regulation” of the real risks.
When the people closest to a technology are also its most careful critics, the responsible move isn’t to dismiss them — it’s to build as if they might be right.
The risks that are already here
Extinction scenarios dominate headlines, but most of the harm we design against is nearer, more mundane, and already measurable. Being honest about AI means naming these plainly:
- Deception and fraud. Engineering firm Arup lost US$25.6 million to a deepfake video call impersonating its executives. Detected deepfake incidents rose roughly tenfold from 2022 to 2023.
- Vulnerable users. A 2025 Common Sense Media survey found 72% of US teens have used an AI companion. In September 2025 the FTC opened an inquiry into companion chatbots and youth safety after lawsuits against major AI firms.
- Cognitive and economic effects. An early, small MIT Media Lab preprint (“Your Brain on ChatGPT”, 2025) reported weaker recall among heavy LLM users — preliminary, but worth watching. On jobs, the IMF estimates ~40% of global employment is exposed to AI, while the WEF projects a net gain of 78 million jobs by 2030 alongside heavy displacement. Both can be true.
- Environmental cost. The IEA puts 2024 data-centre electricity near 415 TWh (~1.5% of world supply), projected to roughly double by 2030. Per-query water estimates have fallen sharply as methods improved — from headline figures for GPT-3 training to Google’s ~0.26 mL for a median Gemini text prompt — but the aggregate footprint is real and growing.
How we choose to build
Naming risks is easy. The harder question is what changes in the work. These are the principles we actually hold ourselves to — the same ones we bring to the teams we work with.
Scope narrow, and keep a human on the irreversible
The safest agent is a specific one. A tightly-scoped system that completes one verifiable workflow is more controllable, more testable and more honest than a general one that half-does ten. On anything irreversible or high-stakes — money moving, data being deleted, a message going to a real customer — a human approves the action, not just reviews the output afterward. We wrote about this discipline in our guide to agentic AI use cases across industries.
Instrument everything, and evaluate against reality
You cannot claim a system is safe if you cannot see what it did. Every tool call and decision should be logged, so behaviour can be audited, debugged and improved. “It felt fine in the demo” is not evaluation; a test set of real cases, measured for task success and failure modes, is. Observability isn’t a nice-to-have — it is the precondition for responsibility.
Be honest about what it is — and what it isn’t
We don’t ship AI that pretends to be human, and we say plainly when an output is model-generated and where it can be wrong. The connectome milestones above are a useful humility check: mapping 166,000 neurons is a staggering achievement, and we still don’t fully understand how that wiring produces behaviour. If we can’t yet explain a fly’s brain, marketing any AI as “conscious” or infallible is dishonest. Measured claims are a safety feature.
Match the pace to the stakes
Speed is not automatically progress. For low-stakes internal tooling, iterate fast. For anything touching health, finance, minors or public trust, the responsible cadence is slower — more evaluation, more guardrails, more human oversight before scale. Building responsibly sometimes means shipping later, or not shipping a feature at all.
The bottom line
Being an AI company that talks openly about AI risk isn’t a contradiction — it’s the job. The technology is genuinely powerful and genuinely double-edged, and the honest position is to hold both facts at once: build the useful thing, and build it as though the careful voices in the field are worth listening to. That’s the standard we set for our own products and the work we do with clients. If you’re weighing where AI fits in your organisation and want a partner who will tell you the risks as clearly as the rewards, let’s talk.
More from across our network
This post is part of a wider conversation we’re having across our ventures about AI, its risks, and staying human in the middle of it:
- The real risks of AI (Netavon) — why telcos, governments and large enterprises must lead on AI safety, observability and governance.
- The AI risk we can’t ignore (yprateek) — a founder’s personal take on why the AI-risk conversation deserves our attention.
- AI and childhood (MySleepyTale) — how AI is affecting kids, and why human-crafted stories still matter.
- AI and the future of human art (Kalacube) — income, copyright and displacement, and how to support real artists.
- Unplug in Parvati Valley (StonedAge) — a digital detox and going back to nature when the feed gets too loud.
- Living well in the age of AI (Musée Living) — staying present and exploring real human experience.
- Reclaiming attention in the age of AI (DeRamaul) — yoga, mindfulness and taking your focus back.
A note on sources. Every figure above links to a primary or reputable source. Estimates of catastrophic risk (for example, Alex Turner’s “one-in-three”) are individuals’ personal judgements, not measurements, and are labelled as such. Early findings, such as the MIT “Your Brain on ChatGPT” preprint, are flagged as preliminary.