Imagine a logistics company whose software has found a better route. Its lorries cover fewer miles and burn less diesel. The saving appears in the fuel bill and, with proper accounting, in its reported emissions. Now suppose the company seeks a performance payment for the routing programme, or carbon credits. How much of the reduction did the programme cause? Against which baseline? Under whose rules? The routing algorithm has no particular reason to know.

Better routing, fewer breakdowns and more efficient charging can reduce emissions. The accountability gap opens when an operational estimate is used to support a claim that requires a different standard of proof. Questions about the vehicle’s condition, the comparison being made and the rules for accepting a result then become decisive.

“Generic AI” here refers to models applied to transport operations whose main job is prediction or optimisation. Some are highly specialised in logistics or engineering; the term describes their use as a general answer to a carbon-accountability problem. This is a distinction about deployment, rather than a recognised technical class. Such models can form part of an environmental accounting system. On their own, they supply only some of what it needs.

Existing transport AI already pursues environmental gains. Google Maps, for example, uses AI to recommend routes expected to consume less fuel or energy. [1] The harder question is how an estimated saving becomes an attributable result that another organisation can inspect, accept and, where appropriate, pay for. That is where the case for Carbon AI begins.

The vehicle keeps changing

A vehicle’s certified specification describes its performance under prescribed test conditions. Its working life supplies a less orderly experiment. In its first report on real-world CO₂ emissions, published in March 2024, the European Commission found that petrol cars registered in 2021 emitted, on average, 23.7% more than their laboratory values suggested. The corresponding gap for diesel cars was 18.1%. For plug-in hybrids, real-world emissions averaged 3.5 times the type-approval figure. These were early results from a limited dataset, not adjustment factors for every car. [2]

Traffic, temperature, auxiliary equipment and driving behaviour help explain the difference; charging habits matter particularly for plug-in hybrids. The findings concern the gap between testing and use, rather than deterioration alone. A known vehicle specification does not describe an unknown pattern of use.

Condition adds another set of variables. Low tyre pressure and engine faults can increase fuel use. [3] A useful system must identify which changes matter and what action is justified. After recommending a repair, it should check whether the expected benefit arrived and endured.

For combustion CO₂, reliable fuel consumption and appropriate fuel characteristics provide a sound basis for calculation, as the IPCC’s mobile-combustion guidance sets out. [4] Condition data helps explain why fuel was used and how to reduce it. More modelled variables do not necessarily make the total more accurate.

A fuel record also cannot establish that a diesel particulate filter remains effective. Such filters can substantially reduce particle emissions; assessing their condition requires appropriate inspection, diagnostics or measurement. [5] A vehicle may look satisfactory on a fuel dashboard while a pollution-control fault goes unresolved. Air-quality compliance and greenhouse-gas accounting need separate results.

Electric vehicles shift the boundary again. They have no tailpipe CO₂, but an assessment including electricity generation must use their energy consumption and the emissions factor required by the chosen method. The calculation follows the question being asked.

The missing comparison

Suppose a fleet uses 8% less fuel per kilometre after driver training and maintenance. It is tempting to award the programme an 8% emissions reduction. Yet the fleet might also be carrying lighter loads or serving shorter, flatter routes. It might travel farther overall. Lower emissions per kilometre can coexist with higher total emissions.

Attributing a reduction to an intervention requires an estimate of what would have happened without the intervention. That counterfactual cannot be read from a sensor. It has to be constructed under an explicit method, using comparable operating conditions and a defensible baseline. A laboratory rating, a pre-repair reading and a crediting baseline answer different questions. Choosing the most favourable one is not an accounting method.

Baselines also need discipline over time. If an algorithm keeps adjusting the comparison to whatever the vehicle does next, it may absorb the very improvement it is meant to measure. Or it may turn deterioration into a more generous starting point. Continuous monitoring should improve the assessment without silently rewriting the terms of the claim.

This is a practical problem for incentives. A driver should not receive a reward merely because the weather improved. An operator who pays for a useful retrofit should be able to show what it achieved. Poor attribution can both reward changes that would have happened anyway and leave worthwhile interventions uncompensated. Better prediction alone does not resolve either error.

Who accepts the claim

Regulation is bringing more of the vehicle’s working life into view. Euro 7 provides for on-board monitoring and an environmental vehicle passport, alongside emissions and battery-durability requirements. [6] Such provisions improve the information available for oversight. They do not turn a vehicle record into an entitlement to carbon credits.

Consider the records needed for a single intervention. The workshop knows what was repaired. The telematics provider records operation. An emissions calculation applies a method to those data. A verifier or authority must establish that the records concern the same vehicle and period, that the comparison is valid and that the claimed use is permitted. For small reductions spread across thousands of vehicles, the cost of reconstructing that chain can determine whether an incentive scheme is viable. Merely assuming the records agree makes it easy to overclaim. An accountability system should connect existing records and request further inspection where a particular decision needs it.

Credit eligibility imposes further tests. The Integrity Council for the Voluntary Carbon Market’s quality principles include additionality, robust quantification, independent validation and verification, and avoidance of double counting. [7] The applicable crediting programme still determines what qualifies. A repair that lowers emissions may support an incentive under one scheme without qualifying for a tradable credit under another. Useful reductions need not all become credits; they do need to be described honestly.

The same applies to tokenization. A token can help administer an asset once the conditions for that asset are satisfied. It cannot supply a missing baseline, establish additionality or confer a regulator’s approval. Automating a weak claim makes it easier to circulate.

What Carbon AI must do

The useful case for Carbon AI is to organise intelligence around these obligations. It should connect the vehicle’s changing physical state with emissions methods, intervention assessment and the rules governing the result. That means working with data of known origin, validated calculations and explicit uncertainty. Another party must be able to examine how the conclusion was reached.

Machine learning can help detect deterioration, estimate a missing variable or identify a promising intervention. A digital twin can organise the changing state of a vehicle. Neither turns an inference into a measurement. When the data cannot support a decision, the useful output may be a request for an inspection rather than a more confident estimate. The system must also retain the version of the model and method used, so a later update does not make an earlier claim impossible to reproduce.

Some of this work benefits from machine learning; much of it depends on sound accounting and reliable records. This places Carbon AI within a wider requirement for transportation decarbonization infrastructure. The calculation must remain connected to the action it assesses and to any resulting payment, compliance decision or asset. Authorities, verifiers and registries retain their responsibilities. Software should make those responsibilities easier to discharge, including correcting a claim when new information undermines it.

Existing AI providers can build this capability, just as a product labelled Carbon AI can fail to deliver it. The distinction rests on the decisions a system can support and the scrutiny its results can survive.

When a vehicle changes, can the system explain the environmental consequence, establish a fair comparison and show why a particular payment or credit is justified? Until those questions are answered, the fuel saving and the proposed entitlement remain separate matters. The first may be visible on a dashboard. The second needs an account that survives examination.

Sources

1. Google — AI and fuel-efficient routing, October 2023

2. European Commission — First real-world CO₂ report, March 2024

3. US Department of Energy and EPA — Keeping Your Vehicle in Shape

4. IPCC — 2006 Guidelines, Volume 2, Chapter 3, Mobile Combustion

5. US EPA — Diesel Particulate Filter General Information

6. European Union — Regulation 2024/1257, Euro 7

7. ICVCM — The Core Carbon Principles