2026 has been a year of two milestones that would each have defined a decade. Scientists published the complete wiring diagram of a whole brain — the male fruit-fly connectome, roughly 166,000 neurons and 125 million connections, mapped by Google Research, Janelia and Cambridge in Cell, following the first full female fly-brain map in Nature in 2024. In the same window, AI systems began generating entire playable worlds from a single prompt, and senior figures across the leading labs publicly argued for pacing the race toward more powerful AI. We can now read a biological intelligence and write synthetic worlds at once — and the people closest to the technology are the ones asking us to slow down.
That is a striking place to stand. But for a telecom operator, a government department, or a large enterprise, the useful question is not whether AI is astonishing. It plainly is. The question is narrower and more practical: what are the real risks, which of them land on us, and what does it take to run this technology responsibly? This is written for the organisations that cannot treat AI as an experiment — because when their systems fail, other people feel it.
The real risks, named honestly
It helps to separate the risks that are already here from the ones that are still contested. Both matter; they just call for different responses.
Fraud and deepfakes — already operational
This is the risk large organisations feel first. In 2024, engineering firm Arup lost about US$25.6 million when an employee was tricked by a deepfake video call impersonating the company’s CFO and colleagues. Identity-verification firm Sumsub reported a tenfold global surge in deepfake incidents between 2022 and 2023. Deloitte projects US generative-AI-enabled fraud losses rising from roughly $12.3 billion in 2023 to $40 billion by 2027. For any organisation that moves money, approves access, or acts on a voice or a video, this is not speculative — it is a live control gap.
Autonomy without oversight
A chatbot suggests; an agent acts. As enterprises move from AI that answers questions to AI that calls tools, moves data and triggers workflows, the failure modes change. An agent can choose the wrong tool, loop on a task, act on stale data, or degrade quietly — and traditional monitoring, which tells you a service is up, will not catch any of it. OpenAI itself has acknowledged that safety safeguards can become less reliable over long interactions. The lesson is not to avoid agentic AI; it is that autonomy raises the bar for what you must be able to see, explain and stop.
Loss of control — the contested, longer-horizon risk
The existential end of the debate is real, but it deserves careful language. In 2023, hundreds of researchers and executives — including Sam Altman, Dario Amodei, Geoffrey Hinton and Yoshua Bengio — signed the Center for AI Safety’s one-sentence statement that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Testifying to the US Senate, Bengio warned that “none of the current advanced AI systems are demonstrably safe against the risk of loss of control to a misaligned AI.” Hinton, on 60 Minutes, said he could not see a path that guarantees safety. Individual probability estimates vary widely and are personal judgements, not measurements — a recent Guardian opinion column by a former Google DeepMind researcher put his own guess of an AI takeover at roughly one-in-three, while others place it far lower. Treat these as informed opinion. The signal that matters for a serious organisation is simpler: the people building the frontier are not claiming it is safe by default.
Cognitive, economic and environmental costs
Three quieter risks round out the picture. On cognition, an early MIT Media Lab study — preliminary, and small — found that people who leaned on a language model to write showed weaker neural engagement and poorer recall of their own work; the authors call it “cognitive debt.” On work, the IMF estimates about 40% of global employment is exposed to AI, and the World Economic Forum projects 170 million new jobs against 92 million displaced by 2030 — transformation to be governed, not a headcount cut to celebrate. On energy, the IEA estimates data-centre electricity use of roughly 415 TWh in 2024, potentially near doubling to about 945 TWh by 2030, with AI the main driver. None of these is a reason to stop. Each is a reason to run AI deliberately.
The people closest to the frontier are not claiming AI is safe by default. For an organisation that runs critical systems, that is the whole brief: assume nothing, verify everything, and never let a system take an irreversible action you cannot see, explain or stop.
Why large organisations have to lead
It is tempting to wait for regulation to settle the question. That is a mistake, for two reasons. First, the rules are unsettled: Canada’s attempt at a comprehensive federal AI statute, the Artificial Intelligence and Data Act, died on prorogation in early 2025, leaving a patchwork. India, likewise, is still shaping its approach. Waiting for a law means running ungoverned AI in the meantime. Second, and more fundamental: telecom operators, government bodies and large enterprises run the infrastructure everyone else depends on. Scale is responsibility. When a bank, a carrier or a ministry deploys AI badly, the harm is not contained to a product demo — it reaches citizens and customers.
The labs themselves point the same way. Demis Hassabis, who leads Google DeepMind, called at the 2026 AI Impact Summit in Delhi for research into AI threats “to be done urgently,” for “smart regulation” of “the real risks,” and for “strong safeguards … to protect against the gravest dangers.” Leadership on safety is not a brake on AI transformation. Done properly, it is the thing that makes transformation possible at all — because no serious organisation will put AI into the core of its operations until it can trust what that AI does.
What responsible AI actually looks like in production
Responsible AI is not a policy document. It is a set of engineering disciplines built into the system from day one. At NETAVON, coming from a telecom operations background, we treat it as the same problem we already solve for networks that cannot go down — see everything, then steer everything — applied to AI.
- Governance: clear ownership of every AI use case — who is accountable, what data it may touch, what it is allowed to do, and what it must never do.
- Observability: traces of every agent run — the steps, tool calls, inputs, outputs, cost and quality — so a bad result can be traced to its cause and a degrading system surfaces before it fails.
- Orchestration and guardrails: a control layer that decides what runs, under which limits, with retries and graceful failure — guardrails built into the flow, not bolted on after an incident.
- Human-in-the-loop on irreversible actions: anything that moves money, changes infrastructure, or cannot be undone routes to a person. Autonomy is earned, scope by scope, once the trail is trusted.
- Audit trails and data residency: a defensible record of what ran, on what data, and why — essential for accountability and for Indian enterprise and government contexts.
None of this is exotic. It is the same reason a carrier network stays up while individual elements fail: full visibility plus real-time steering. We have written separately about the two halves of that discipline — how observability keeps agentic AI reliable, and how orchestration turns scattered models and agents into one governed system.
We built an interactive showcase for the two disciplines that underpin responsible AI — observability (telemetry off vs on) and orchestration (self-healing reroute, QoS), grounded in the fully-mapped fruit-fly connectome.
Open the interactive showcase →From risk to responsible transformation
The honest framing is neither doom nor hype. AI is powerful, its risks are real, and most of the ones that will actually hit a large organisation this year are governable with discipline rather than luck. The organisations that win the next few years will not be the ones with the most pilots, nor the ones that sat out AI waiting for certainty. They will be the ones that led on safety — that could put a use case live, prove it, govern it, and repeat — because they built the observability, orchestration and human oversight to earn trust at every step.
That is the work NETAVON does: responsible AI and AI transformation for government, telco, satellite, NTN and large enterprise across India — starting from operations, with safety built in, not bolted on.
If you are putting AI into the core of your operations and safety cannot be an afterthought, that is exactly the work we do. Tell us the use case and we’ll tell you how we’d govern it.
Let’s talk — book a callRelated from NETAVON
- What a fly brain teaches telecom networks — The observability and orchestration disciplines behind responsible AI, seen through the fully-mapped connectome.
- Observability & orchestration showcase — An interactive look at seeing and steering a large distributed system in real time.
More from across our network
- Building AI responsibly — How an AI company thinks about building the technology responsibly, from the inside.
- The AI risk we can’t ignore — A founder’s personal perspective on the AI-risk conversation.
- AI and childhood — What AI means for kids — and why human-crafted stories still matter.
- AI and the future of human art — Income, copyright and displacement — and supporting real artists.
- Unplug in the Parvati Valley — A digital detox and a case for stepping back into nature.
- Living well in the age of AI — Staying present and living real human experience as AI advances.
- Reclaiming attention in the age of AI — Yoga, mindfulness and reclaiming focus in a distracted world.
Sources: Center for AI Safety statement on AI risk (safe.ai, 2023); Yoshua Bengio, US Senate testimony (2023); Geoffrey Hinton, CBS 60 Minutes (2023); Demis Hassabis, AI Impact Summit Delhi, via Malay Mail (2026); Arup deepfake fraud, Fortune / CFO Dive (2024); Sumsub deepfake incident report (2023); Deloitte, gen-AI fraud projections (2024); MIT Media Lab, “Your Brain on ChatGPT” preprint (2025); IMF, AI and the global economy (2024); World Economic Forum, Future of Jobs 2025; IEA, Energy and AI (2025); fruit-fly connectome, Nature (2024) and Cell / Google Research, Janelia & Cambridge (2026). Probability-of-catastrophe estimates are individual opinions, not measurements, and are presented as such.