How Does AI Affect your ESG Goals?

Summary

AI now sits on both sides of the ESG ledger, and most large enterprises have to report on which side wins.

In this session, host Richie Cotton puts three practitioners to work on a question that has moved from academic to mandatory: what does artificial intelligence do to a company's environmental, social, and governance goals? Saskia van Gendt, chief sustainability officer at supply-chain platform Blue Yonder, tracks the running balance between AI's efficiency gains and its electricity appetite. Padmini Murthy, chief marketing officer at RiskOps AI, brings the governance view, where she says security breaches and lost digital trust keep her customers awake at night. James Brusseau, a philosophy professor at Pace University who runs AI ethics audits, reframes the social questions as genuine dilemmas with no clean answer.

The conversation moves through the hard numbers behind data-center power demand, the renewable and nuclear bets meant to cover it, the social fallout from deepfakes and machine-scale bias, and the accelerate-versus-slow-down argument now splitting the field. It ends on practical ground: how a company audits its AI systems, who stays accountable, and where to start. The panel disagrees often, which is the point. Watch the full recording for the surgery thought experiment, the fairness-versus-privacy trade-off, and the one framework Brusseau says everything else derives from.

Key Takeaways

• Data centers consumed roughly 450 terawatt-hours of electricity in 2025, and that figure is on track to double by the early 2030s as AI's share of capacity climbs from about 20% to 40%.
• Only 6% of organizations running AI agents have budget set aside to secure them, even as breaches happen daily.
• An estimated 92 million jobs could be displaced globally by 2030.
• AI can route a heavy compute task to a region where the grid is cleaner and demand is lower, turning AI into a tool for its own efficiency.
• In narrow medical domains, AI already out-diagnoses humans, which forces a harder question than "should we add oversight?" — adding a human can introduce error.
• Fairness and privacy can pull against each other: checking whether a hiring model treats groups equally requires collecting the sensitive data that privacy rules tell you to leave alone.
• The EU's Ethics Guidelines for Trustworthy AI (2019) is the panel's recommended starting point for AI impact assessments.
• The practical first move for any company is a clean audit of every AI system it runs, with named owners for each decision.

Deep Dives

The disasters already showing up

Cotton opens by asking each guest for the worst ESG outcome they have seen from AI. Murthy goes first and names two. The first is employment. She cites a projection that "by 2030, they're gonna be about 92,000,000 jobs that are gonna be displaced globally," and ties it to a personal worry about a generation steered into computer science and data science with no clear landing spot. The second is security. As anyone can now spin up an agent to run their own tasks, those agents reach into large stores of data, and the spending to protect them has not kept pace. "Only 6% of them have the budget for securing those agents," she says, "and the breaches are happening every single day."

Van Gendt declines the word "disaster." After 25 years in sustainability, she frames AI as a running balance sheet she and her peers monitor in real time: carbon and energy savings on one side, data-center expansion and electricity demand on the other. She sees real upside in supply chains, where AI sharpens inventory planning and cuts the overproduction that drives waste, improves route and load building, and pivots sourcing around climate risk in real time.

Brusseau, the philosopher, recasts the framing entirely. The representative cases he meets sit on the social side, and the clearest is health care. With enough early data, he says, AI can predict the diseases and even the psychological and educational strengths a person will develop. That power is already being used in places with centralized governments to "direct people's lives." The benefit is real and measurable; so is the cost. The question he leaves on the table: how much individual freedom and privacy will we trade for collective gain?

Powering AI's energy appetite

The electricity number anchors the environmental discussion. Van Gendt notes that data centers used about 450 terawatt-hours in 2025 and are expected to roughly double, while AI's share of data-center capacity grows from around 20% toward 40%. Her job, she says, is to "make sure that we have the electricity where it's needed to power that growth and that, hopefully, that electricity can be supplied with clean energy."

She is more optimistic than the headline suggests. Renewable capacity keeps rising even as US incentives fade, and she points to a milestone: "I think solar energy surpassed, excuse me, surpassed coal energy for the first time, coal production." Storage excites her most. Because US grids are built for peak load but average around 50% use, distributed batteries can capture what is already there. Someone described it to her as storing "the electrons of the future." Add new nuclear, which the hyperscalers are funding, and the major data-center operators have, she notes, "all committed to net zero goals for decarbonizing their operations."

Murthy is less sanguine and says so plainly. Compute demand at current scale points toward shortage: "there's gonna be a crisis because the the way I see companies using compute today... every query I do on an LLM is using high compute power." She argues the fix has to reach policy level, not just company level, and points to the EU AI Act as one lever. One application both find genuinely clever is intelligent routing: send a heavy AI task to a region where the grid is cleaner and off-peak. As van Gendt puts it, "it's almost like AI for AI efficiency."

The social cost: deepfakes, bias, and AI companionship

Murthy's social concern is erosion of trust. "One of the biggest things that worries me is the whole concept of digital trust, which we are losing to things like deep fakes," she says. Alongside that sit machine-scale discrimination from algorithmic bias and agentic systems that act with no human in the loop. These stop being technical problems the moment a model's wrong answer decides who gets hired or who gets a loan.

Brusseau pushes into stranger territory: AI companionship and what it does to the metrics we use for human well-being. AI friends may well make people happier on every surface measure, partly because, as he notes, there is "almost certainly less depression since their friends are always so sycophantic." But he doubts our industrial-age definitions of happiness and friendship will survive contact with the technology. We talk to pets and plants without calling them friends; AI will force us to define what friendship actually is. "Is AI really a friend? What counts as a friend?" He admits we lack both the answers and the framing.

Van Gendt keeps the thread tied to people. The reason she works on climate and pollution, she says, is human benefit, and she wants human-centered design, transparency, and oversight to carry into AI. In supply-chain work that means keeping a person able to review, override, and escalate on higher-risk calls: the AI produces "a recommendation, not necessarily an action in all cases." Murthy agrees the human has to sit at every layer, from model building to deployment to regulation.

Accelerate or slow down — and who stays in the loop

Brusseau names the fault line under most AI debate: acceleration versus precaution. The acceleration camp argues that the cure for AI's problems is more AI, faster — better models, satellites collecting solar power in space, quantum computing that makes the energy problem evaporate. He reads Google's spending on expensive free tools as a signal it may be closer to quantum computing than outsiders think. The precaution camp wants to slow consumption until renewables and storage catch up. ESG professionals, he suspects, skew precautionary, and the split maps loosely onto America versus Europe.

The human-in-the-loop consensus gets a sharp challenge from him. Everyone Cotton has interviewed wants a human accountable and watching, while AI vendors push to remove humans as bottlenecks. Brusseau sides, narrowly, with the data: "even when you put the human in the loop, you introduce error. The AI is just so much better than humans that actually adding that level of oversight makes it worse, not better." He offers his students a thought experiment — a surgery with "a ninety five percent chance" of success under AI versus "an eighty five percent chance" under a human who can explain what is happening. How much survival probability would you trade to understand what is being done to your body? Some students take the numbers; some keep the human. Murthy lands on explainability as the deciding factor: outside narrow domains like health care, people need to know how a decision was reached, and that demand only grows as AI scales.

Governance, auditing, and where companies should start

When Cotton turns to implementation, Murthy gives the most concrete answer. "The first step here is to do a clean audit of everything that you're building in your AI system," she says, paired with assigning roles so everyone from developer to regulator owns part of the decision. Add compliance and policy teams early, run periodic data-bias reports where demographic data exists, and "train your teams to detect drifts, to detect hallucinations." RiskOps AI, she notes, is in the business of closing that audit and risk-assessment gap.

Brusseau sharpens the hardest governance choice into a single decision a company actually has to make. To know whether a hiring model treats groups equally, you have to know each candidate's sensitive attributes — and collecting that data runs straight into privacy. "Do you want to be fair, or do you want to respect user's privacy? You have to choose one or the other," he says. These are new dilemmas the industrial economy never posed, and a company's ESG profile now depends on how it answers them.

Van Gendt's advice is integration over isolation: mandates set at the top, departmental goals beneath them, and technology giving teams the real-time visibility to compare a low-carbon route against a faster or cheaper one. Asked for one framework to anchor AI impact assessments, Brusseau does not hesitate: the EU's "Ethics Guidelines for Trustworthy AI" from 2019. "That's the standard. Everything derives from that. That's like The Beatles." On where to run a model, van Gendt closes with governance as policy constraints — run a workload only where grid carbon intensity sits below a set threshold, or only inside a specified jurisdiction — so efficiency and compliance hold together.