NiqCurry.com

Category: Human Behaviour

  • Who Chooses Now? AI and the Gap Between Values and Behaviour

    TL;DR

    • Where options appear matters more than labels. In a 2024 UK trial, listing lower-carbon dishes first cut the carbon in each order by 12%; carbon labels and a meat tax made no significant difference. (Section 1)
    • AI now decides where options appear. An AI agent has no gap between values and action, so whoever sets its objectives largely decides what customers buy. (Section 2)
    • The legal risk is already live. EU rules on green claims apply from 27 September 2026, UK fines reach 10% of global turnover, and claims written by AI count. (Section 3)
    • Defaults decide who pays. A green default that costs more falls hardest on poorer households, and brands will need sustainability data that agents can read. (Section 3)
    • One owner, one test. Give a named person responsibility for what your AI optimises for, and run six questions before launch. (Sections 4 and 5)

    The small decisions AI is starting to make for us

    In January 2026, Google’s chief executive Sundar Pichai used the National Retail Federation’s annual show in New York to launch the Universal Commerce Protocol, an open standard that lets AI agents find, compare and buy products on a shopper’s behalf. It was built with Shopify, Etsy and Wayfair, among others, and endorsed by more than 20 companies including Visa and Mastercard. Grocers are moving the same way: Albertsons’ AI assistant now offers “rapid restock” of weekly staples across Safeway, Vons and its other chains. Systems like these are starting to make many of the small decisions that determine how sustainably people live: which dishes Just Eat lists first, which staples Albertsons’ assistant reorders, which route Google Maps recommends.

    Recent research shows that the order in which options appear can shape choices more than information or price. In a 2024 trial with 4,008 UK consumers using a simulated food-delivery app, listing lower-carbon dishes first cut the carbon footprint of the average order by 12%. Carbon labels and a visible meat tax, tested in the same trial, made no significant difference.

    Ordering, ranking and pre-selecting are exactly the jobs organisations are now handing to AI. An AI agent has no gap between values and action. It does what it was set up to do. As more purchases pass through these systems, the objectives they are given will do more to decide what customers buy than customers’ own convictions, and in most organisations nobody has yet been made responsible for those objectives.

    The most urgent consequence is legal. From 27 September 2026, EU law bans unsupported generic green claims, including those written by AI, and UK regulators can already fine up to 10% of global turnover for misleading ones.

    1. Why don’t people buy what they say they value?

    People’s stated support for sustainability runs well ahead of what they buy. In the US and Canada, 76% of shoppers call themselves conscious consumers, yet values-driven purchases make up 38% of their buying. The usual explanation is that people don’t mean what they say. But products marketed as sustainable hold 25.4% of US consumer goods sales against 36.8% in the UK and 42% in Germany, in countries with broadly similar attitudes. The difference lies in the buying environment, not the buyers.

    The clearest recent evidence of how much the environment matters comes from a 2024 study led by Cambridge behavioural economists. Researchers built a platform that mimicked food-delivery apps such as Just Eat and Deliveroo, and asked 4,008 UK adults to order a meal. They compared three approaches. One added a carbon label to each dish. One added a visible tax on meat. One simply changed the order of the menu so that lower-carbon restaurants and dishes appeared first.

    Only the reordering made a significant difference. It cut the carbon footprint of the average order by 12% and reduced high-carbon meal choices by 13%. People who saw the reordered menu also chose meals with fewer calories and more nutritional value, and reported being more satisfied with what they ordered. The researchers estimate that if Just Eat alone made the change, it would avoid 78,000 tonnes of CO₂ a year in the UK, at a cost of £1.28 to £3.85 per tonne. Information and price signals, the tools organisations usually reach for, did less than the order in which options appeared.

    Field research points the same way. A 2025 study that made vegetarian options the default in a real food setting found vegetarian sales more than doubled and meat sales fell to under a third of their previous level.

    Two cautions keep this in proportion. First, information can backfire. A 2025 Irish study found that relative flight labels (“23% less emissions than average”) led some people to choose higher-emitting flights, while absolute figures alone changed nothing. Second, not every digital nudge works. A 2024 meta-analysis of 67 studies of digital green nudges, after correcting for the tendency to publish only positive results, found no significant average effect. The lesson is not that design changes always work. It is that ordering and defaults at the point of decision can have large effects, that lighter-touch prompts often have none, and that the only way to know is to test.

    2. What happens when AI decides which option comes first?

    The food-delivery study changed a menu by hand. In practice, the order of a menu, a search results page or a product list is increasingly decided by algorithms, and the decision to buy is starting to be made by AI agents acting for customers. The features of a purchase that the evidence shows matter most (what appears first, what is pre-selected, how a comparison is described) are moving from designers’ hands into AI systems.

    Customers are handing over the routine and keeping the significant. Only 11% of US consumers would let AI make purchase decisions even in low-stakes categories. Yet groceries (41%) and household supplies (31%) are the categories they most readily delegate, and consumers expect agents to make 15% of their purchases within five years, up from 9% a year earlier. Routine, repeated purchases are where everyday environmental impact accumulates, and where ordering and defaults have the most effect.

    This moves the gap between values and behaviour somewhere new. An AI agent has no such gap of its own. It does not get tired, rushed or distracted, and it does not have values that it then fails to act on. It follows the objectives it was given. If those objectives put price and speed first, lower-impact options will lose almost every time, whatever the customer believes. If they include the customer’s stated preferences, the gap could narrow across millions of purchases at once. Since 57% of consumers would let an agent switch brands for better value, whatever the agent counts as value will also decide which brands win.

    AI-controlled choices also differ from a store layout or a sign-up form in ways that raise the stakes:

    • Scale. One setting applies to every customer and every order at once.
    • Repetition. A weekly reorder repeats the same choice hundreds of times, so small biases add up.
    • Invisibility. Customers rarely see the criteria behind a ranking or a recommendation, so they can’t easily notice or correct them.
    • Personalisation. The framing can be tailored to each person, which makes it more effective and harder to audit.

    The effect depends on the purchase. For routine purchases that people delegate, the agent’s objectives effectively are the decision; customers’ values only reach the purchase if someone built them in. For larger purchases that people still make themselves, AI shapes what they see and how options are described. The flight-label research shows that even well-meant framing can push people the wrong way.

    The result is a new and largely unrecognised role. The person who matters most for many customers’ environmental choices is now often a product owner or data scientist setting an algorithm’s objectives. Most organisations have not yet acknowledged that this role shapes customers’ sustainability outcomes, let alone given anyone accountability for it.

    Some companies already show what careful configuration looks like. Tibber, an energy supplier founded in Norway that now also operates in Sweden, Germany and the Netherlands, asks electric-car owners to set just two things in its app: how full the battery should be and when the car is needed. Its smart-charging system then decides when to charge, moving it to the hours when power-exchange prices are lowest. Analysing more than 10,000 German households in the first half of 2026, Tibber found their charging cost less than a quarter of the energy cost of a diesel car per 100 km. In the Netherlands, almost half of its smart-charging sessions in 2025 took place between noon and 5pm, when the cheapest power coincides with the solar peak. Customers keep control of the outcome they care about, and the system optimises everything else.

    The food-delivery study points the same way: one well-chosen ordering decision cut emissions, improved nutrition and left customers more satisfied, at very low cost. AI makes it possible to apply decisions like these consistently, across every customer and every purchase, while still leaving people free to choose otherwise.

    3. Which consequences need action now, and which decide trust over time?

    The shift has consequences across five dimensions: governance, economic distribution, human flourishing, society and the planet. They are not equally pressing. For most organisations, two need action now; the other three set the conditions for whether the approach is trusted and sustained.

    Who is liable when AI makes a green claim?

    The EU’s rules on green claims apply from 27 September 2026, with no transition period, including for products already on shelves and claims already published. Generic terms such as “eco-friendly” or “green” are banned unless the business can demonstrate recognised excellent environmental performance. In the UK, the Competition and Markets Authority can fine up to 10% of global turnover for misleading claims, and its 2026 guidance makes retailers liable for repeating a supplier’s claim simply by offering the product for sale.

    AI multiplies this exposure. A product assistant that summarises supplier descriptions repeats their claims. A model generating copy for a whole catalogue can introduce vague green language across thousands of pages in one run. And staff often don’t check: in the UK, 58% of employees using AI said they had relied on its output without checking it. Meanwhile, the EU AI Act’s high-risk obligations have been delayed to 2027 and 2028. For most organisations, the most immediate legal risk from AI is not AI law. It is consumer protection law applied to what their AI says.

    Who pays for a green default, and who gains from it?

    Green defaults often ask the customer to pay a little more for a collective benefit. A 2025 review found that the same nudge is likely to have a stronger effect on poorer people, who also contribute less to pollution. A default that costs more can therefore fall hardest on those least able to pay. When an AI agent applies that default to millions of households, the unfairness scales with it.

    Value is also moving. Products marketed as sustainable delivered nearly 45% of US consumer goods growth from 2013 to 2025, so the stakes for brands are high. If agents choose between brands, a sustainability credential only counts if the agent can read and verify it. Marketing language won’t register; specific product data will. Influence that used to sit with brands moves towards whoever sets the agent’s criteria.

    Will people trust AI that chooses for them, and does it save more than it costs the planet?

    The remaining three dimensions decide whether customers and the public accept this shift.

    For human flourishing, the question is whether AI extends people’s ability to act on their values or quietly replaces it. People already use AI more than they trust it: 66% use it regularly, 46% are willing to trust it. Delegation stays legitimate only if customers can see what is being chosen for them and change it.

    For society, the issue is the shared information people rely on. 49% of consumers have abandoned products over unclear or misleading sustainability messages, and trust in organisations to handle data is declining. Organisations using AI to describe products are collectively setting the quality of information a whole market depends on.

    For the planet, AI has its own rising footprint. Google’s electricity use grew 37% in 2025 and its water use 34%, and global data centre electricity demand is expected to more than double by 2030. An organisation using AI to promote greener choices without counting the AI’s own footprint leaves itself open to challenge on credibility.

    AI can also reduce impact directly, out of customers’ sight. Picnic, the Dutch app-only grocer now expanding in Germany and France, orders stock from demand forecasts built on every customer order rather than estimates of who might walk into a shop. It reports food waste of 0.72% of what it sells and estimates it wastes around seven times less food than conventional supermarkets. Uses like this, where AI removes waste rather than generating more activity, are the ones most likely to come out ahead once the AI’s own footprint is counted.

    4. How can you tell whether an AI feature will do more good than harm?

    Four principles run across all five dimensions: safety and resilience, justice and fairness, agency and accountability, and long-term and global responsibility. Six questions turn them into a test any organisation can apply before launching AI that shapes customer choices. Here they are applied to one case: an automatic reordering feature like Albertsons’ “rapid restock”, in which an AI assistant reorders a household’s weekly staples.

    QuestionApplied to the reordering agent
    Benefit: what real good could it create?Saves customers time. Set to prefer lower-impact products where price and quality are close, it could narrow the gap between values and behaviour across a large share of everyday spending
    Harm: what could go wrong?Set on price alone, it moves customers away from products they would have chosen. It may repeat unsupported supplier claims
    Distribution: who gains, who loses?A default that costs more falls hardest on lower-income households. Brands without machine-readable credentials lose out
    Agency: who decides, and can people challenge it?Customers must be able to see the agent’s criteria, set their own preferences and reverse any order easily
    Resilience: what if it fails or is misused?An error repeats across thousands of accounts before anyone notices, unless checks catch it
    Time horizon: what about future generations?Small weekly choices across millions of households add up to a measurable environmental effect over years, in either direction

    The value of the test is that it turns a technical question, what the agent should optimise for, into a decision with visible consequences and a named owner.

    5. Who should decide what your AI optimises for?

    One decision matters more than the rest: give a named person responsibility for what the organisation’s AI optimises for when it shapes customer choices. Today that responsibility is usually spread across product, data, marketing and sustainability teams, which in practice means nobody holds it. The objectives of a ranking model or a buying agent are now among the most consequential sustainability decisions an organisation makes, and they need an owner with the authority to set them and the duty to explain them.

    That owner’s first tasks follow from the evidence:

    • Map and write down the objectives. List every place AI ranks, recommends, pre-selects or buys for customers, and record what each system is optimising for. If customers’ stated preferences aren’t among the objectives, the system won’t reflect them.
    • Test before assuming. The food-delivery result and the meta-analysis together show that some design changes have large effects and others none. Run controlled tests on real choices before scaling.
    • Treat every AI-generated claim as a published claim. Check sustainability language against evidence before it goes live, use absolute figures alongside any percentage, and audit existing content, including supplier claims, against the September 2026 rules.
    • Make defaults visible, reversible and fair. Tell customers what has been chosen for them and why, make changing it as easy as accepting it, and check who pays before a default goes live.

    Two further steps matter over the longer term: publishing sustainability credentials as verifiable data that agents can read, and asking AI providers for usage-level emissions figures so the AI’s own footprint can be counted against what it saves.

    The measure that ties this together is the gap between what customers say they want and what they buy through your channels. If AI-assisted channels narrow it without raising complaints, opt-outs or costs for those least able to pay, the approach is working.

    Will the people setting AI’s objectives act on customers’ values?

    The gap between people’s values and their purchases has usually been treated as a weakness in customers. Recent evidence shows it depends heavily on how choices are presented, and AI is now taking over that presentation for a growing share of everyday purchases.

    Well-configured AI could do what years of labelling and awareness campaigns have not: put lower-impact options in front of people at the moment of decision, describe them honestly, and leave people free to choose otherwise. Configured carelessly, the same systems will optimise for price and speed, repeat unsupported claims and add to a footprint nobody is counting.

    The question is no longer only whether customers will act on their values. It is whether the people setting AI’s objectives will.

    The frameworks referred to here are developed further in Actions for Innovation. [Link]

    Sources

    Research

    Company cases

    Market and attitude data

    AI footprint and regulation