Responsible AI by design is the message a group of academic editors put into print this week, and it is aimed squarely at how data science teams currently work. The latest issue of the INFORMS Journal on Computing carries a special issue titled “Responsible AI and Data Science for Social Good”, and its co-editors argue that ethics cannot be a review stage bolted onto a finished system.
The issue was co-edited by Ram Ramesh, PhD, an area editor of the journal and professor of management science and systems at the University at Buffalo School of Management. His summary is blunt. “Responsible AI cannot be achieved by addressing ethical consequences after the systems are built,” Ramesh says. That single sentence is the whole argument for responsible AI by design, and it is the opposite of how most AI governance is staffed and scheduled today.
Across 13 studies the special issue examines judicial systems, education, communication, health care, bias and fairness, and interpretability. One health care paper is the sharpest illustration: a machine learning appointment scheduling system left Black patients waiting roughly 30% longer than non-Black patients, and the fix was not a policy memo. It was a change to what the model was told to optimise. That distinction is what responsible AI by design actually means in practice, and it is what this article unpacks.
Table of contents
- What Researchers Mean by Responsible AI by Design
- Why Bolt-On Ethics Fails and Responsible AI by Design Does Not
- The Scheduling Study That Made the Case Concrete
- How Bias Compounds When Nobody Is Watching
- The Six Domains Responsible AI by Design Has to Cover
- What the Governance Numbers Say About Responsible AI by Design
- Building Responsible AI by Design Into a Real Pipeline
- Responsible AI by Design Questions People Are Asking
- References and Further Reading
What Researchers Mean by Responsible AI by Design
Responsible AI by design is not a new slogan for existing ethics review. It is a claim about where in the build the decisions get made, and it comes with a specific target: the objective function.
The special issue behind the claim
The work appeared in the INFORMS Journal on Computing, volume 38, issue 4, published in September 2026 and reported by the University at Buffalo on 2 September. It gathers 13 peer-reviewed studies under one editorial argument rather than one experiment, which is why it reads as a position on method rather than a single result.
Who signed the argument
Seven academics co-edited the issue: Ram Ramesh at the University at Buffalo, Kaushik Dutta at the University of South Florida Muma College of Business, Ajay Kumar at Emlyon Business School, Zhiling Guo at the University of North Texas G. Brint Ryan College of Business, Martin Bichler at the Technical University of Munich, Dursun Delen at Oklahoma State University, and Paul Brooks at Virginia Commonwealth University. Ramesh is the journal’s area editor for machine learning.
The sentence that carries the weight
“Social goals must be incorporated throughout the AI life cycle, from deciding what data to collect and how models learn to determining what they optimize, how they interact with people and institutions, and how they are governed,” Ramesh says. Read that as a list of five checkpoints, not one. Responsible AI by design means all five, not the last one.
What the editors are not claiming
Responsible AI by design is not an argument that every model needs a fairness term bolted into its loss function, and the editors do not say so. The claim is narrower and harder to dodge. Wherever a system allocates something scarce — an appointment, a place, a score, a queue position — the allocation rule is a design decision that somebody has to own before training starts. Systems that describe the world are a different case from systems that divide it up.
| Detail | Value |
|---|---|
| Publication | INFORMS Journal on Computing, Vol. 38, No. 4 |
| Special issue title | Responsible AI and Data Science for Social Good |
| Studies included | 13 |
| Co-editors | 7, across 7 institutions |
| Domains covered | Judicial, education, communication, health care, bias and fairness, interpretability |
| Lead quoted editor | Ram Ramesh, University at Buffalo |
| Announced | 2 September 2026 |
| DOI | 10.1287/ijoc.2026.ed.v38.n4 |
Why Bolt-On Ethics Fails and Responsible AI by Design Does Not
Most organisations do have an ethics process. The researchers’ complaint is not that it is missing but that it runs too late to change anything that matters.
Review boards inherit finished decisions
By the time a system reaches an ethics review, the training data has been chosen, the label definitions are fixed, the objective is written and the deployment context is set. A reviewer can approve, delay or block. What they cannot do is change what the model was rewarded for. Responsible AI by design moves the decision upstream to the point where it is still cheap.
Predictions are not the product
The researchers make a point that sounds obvious and is routinely ignored: AI systems do far more than generate predictions. They allocate resources, shape opportunities, influence institutional processes and ultimately affect long-term societal outcomes. A scheduler does not predict a wait time — it assigns one. Treating that as a forecasting problem hides the allocation that is actually happening.
Fairness as an objective, not a constraint
“Fairness should not be viewed as an external constraint imposed on otherwise optimal systems but as a core design objective,” Ramesh says. The distinction is technical, not rhetorical. A constraint is something you satisfy at the boundary of a solution you already chose. An objective is something the optimiser trades off from the beginning, which is the whole practical content of responsible AI by design.
| Life-cycle stage | Bolt-on ethics review | Responsible AI by design |
|---|---|---|
| Data collection | Not in scope | Who is represented is a design choice |
| Label definition | Assumed given | Proxy targets challenged before training |
| Objective function | Untouchable by review | Fairness sits inside it |
| Interaction design | Rarely examined | Institutional effects modelled |
| Governance | The entire process | One of five checkpoints |
| Cost of a change | Retraining and rollback | A line in the specification |
| Typical outcome | Approved with caveats | Disparity measured and closed |
The Scheduling Study That Made the Case Concrete
Editorial arguments are easy to nod at. The health care example published in the issue is harder to wave away, because it attaches a number to the problem and a number to the fix.
A 30% wait-time gap
Researchers found that a machine-learning appointment scheduling system had Black patients waiting approximately 30% longer than non-Black patients. Nobody wrote a rule that said so. The system was optimising something reasonable — throughput, utilisation, no-show risk — and the disparity fell out of it.
The fix was in the objective, not the paperwork
The disparity was eliminated by incorporating race into the system’s optimisation objective, while maintaining scheduling efficiency. That is the finding that should change how teams work. The trade-off everyone assumes exists between fairness and performance did not appear here; the efficiency held.
Why removing the attribute would not have worked
The intuitive fix is to strip race out of the model. It fails, because the disparity was produced by the objective interacting with correlated features, not by the attribute being present. Blinding the model removes your ability to measure the gap without removing the gap. Responsible AI by design says the opposite: measure it explicitly, then optimise against it.
What the result does not prove
One system, one setting, one disparity. The study does not establish that fairness is free in general, and responsible AI by design does not require believing that it is. What it does establish is that the trade-off is worth testing rather than assuming. Here the efficiency held and the gap closed, which is at least one counter-example to the reflex that fairness has to be paid for in throughput.
How Bias Compounds When Nobody Is Watching
The researchers are careful to say that bias does not usually arrive as a visible mistake. It develops in less obvious ways, and the mechanisms they list are the ones a code review will never surface.
Historical inequity is already in the data
Training sets record what institutions did, not what they should have done. A model fitted to that history learns the pattern faithfully, which is exactly the problem. Nothing in the pipeline flags it, because by every conventional metric the model is performing well.
Feedback arrives incomplete
You observe outcomes for the people the system selected and almost nothing about the people it did not. Rejected applicants generate no repayment history; unscheduled patients generate no visit record. The model’s view of its own errors is structurally one-sided, which is why offline accuracy keeps looking healthy.
Institutions interact
Disparities also emerge from interactions among different institutions — a scoring system in one organisation feeding an eligibility rule in another, neither of which is unfair in isolation. No single owner sees the composite effect, and no single audit covers it.
The loop closes
Because algorithms influence the data generated by future decisions, those disparities compound over time. This is the same dynamic that makes reinforcement learning systems drift when their reward signal is mis-specified: today’s output becomes tomorrow’s training data. A small gap at launch is not a small gap three years later, and that compounding is the strongest practical case for responsible AI by design.
The Six Domains Responsible AI by Design Has to Cover
The special issue is organised around six areas. They are worth reading as a checklist of where allocation decisions get made by software.
Judicial systems
Risk scores and case-management tools shape bail, sentencing and supervision decisions. The allocation is liberty, and the feedback loop is decades long, which makes responsible AI by design less optional here than anywhere else on this list.
Education
Placement, early-warning and admissions models decide who gets attention and who gets flagged. Errors here compound across a school career rather than a quarter.
Communication
Ranking and moderation systems decide what reaches whom. The optimisation target — engagement, most often — is precisely the design choice the researchers say should be argued about explicitly rather than inherited. Responsible AI by design treats that target as contestable, not as a given of the platform.
Health care
Scheduling, triage and resource-allocation models are where the 30% wait-time finding sits. Health care is also where measurement is best, which is why the disparities get found there first.
Bias and fairness
The methodological core: how to define a disparity, how to measure it, and how to put it into an objective without breaking the system. Responsible AI by design lives or dies on this being tractable.
Interpretability
If you cannot explain what a model optimised, you cannot argue about whether it should have. Interpretability is what makes the other five domains auditable rather than merely assertable.
What the Governance Numbers Say About Responsible AI by Design
Research consensus is one thing; organisational behaviour is another. The 2026 AI Index from Stanford HAI gives a picture of an ecosystem that has adopted the paperwork of responsible AI faster than the practice.
Policies are nearly universal now
The share of organisations with no responsible AI policy at all fell to 11%, down from 24% a year earlier, against a backdrop where 88% of organisations report using AI in at least one business function. On paper, the governance gap is closing.
Incidents rose anyway
Documented AI incidents reached 362 in 2025, up from 233 in 2024 — a rise of roughly 55% in a single year. Policy adoption and incident growth moved in the same direction at the same time, which is what you would expect if the policies sit downstream of the design decisions rather than inside them.
Transparency moved backwards
The Foundation Model Transparency Index average fell from 58 to 40 out of 100 between 2024 and 2025, and 80 of the 95 notable models released in 2025 shipped without published training code. Responsible AI by design assumes you can inspect what a system was trained on and optimised for; that assumption is getting harder to satisfy from the outside.
What the gap means for a build team
The inverse of that 11% is an 89% that does have a policy, and it has not stopped incidents rising by roughly 55% in a year. That is not an argument against writing policy. It is an argument that policy and design are different controls doing different jobs, and that only one of them reaches the objective function. Responsible AI by design is the second control, and these numbers suggest most organisations have bought the first and assumed it covered both.
Which rulebooks organisations actually cite
Asked what shapes their responsible AI practice, organisations named the GDPR most often at 60%, down from 65% the previous year, followed by ISO/IEC 42001 at 36% and the NIST AI Risk Management Framework at 33%. Two of those three are AI-specific and both are minority answers, which tells you how young this discipline still is.
The regulatory clock is not as tight as it was
European deadlines moved in 2026. The Digital Omnibus on AI deferred the AI Act’s standalone high-risk obligations from 2 August 2026 to 2 December 2027, and product-embedded high-risk systems from August 2027 to August 2028. Article 50 transparency duties still landed on 2 August 2026. The extra time is real, and it is exactly the window in which upstream design work is cheapest to do.
| Obligation | Original date | Current date |
|---|---|---|
| Prohibited AI practices | 2 Feb 2025 | In force |
| General-purpose AI model duties | 2 Aug 2025 | In force |
| Article 50 transparency | 2 Aug 2026 | Unchanged |
| Annex III standalone high-risk | 2 Aug 2026 | 2 Dec 2027 |
| Annex I product-embedded high-risk | 2 Aug 2027 | 2 Aug 2028 |
Building Responsible AI by Design Into a Real Pipeline
None of this is useful as a principle. It is useful as a set of artefacts your team produces before the first training run, and the five life-cycle checkpoints Ramesh names map cleanly onto them.
Write down what you are optimising
Not the metric name — the sentence. “We are maximising clinic throughput subject to no-show risk” is a claim about who waits longer, and it should be legible to someone who is not an engineer. If nobody on the team can state it in one sentence, that is your first finding, and responsible AI by design has nowhere to start until it is written down.
Measure the disparity before you ship
You cannot manage a gap you never computed. Pick the groups that matter for your context, compute the outcome difference, and record it as a launch metric alongside accuracy. The scheduling study only exists because somebody measured a wait time by patient group. Responsible AI by design is mostly this step, done early enough to still be cheap.
Put fairness inside the objective
This is the step almost everyone skips. Adding a fairness term to the loss or the constraint set is a modelling task, and it belongs to whoever owns ML model development, not to a governance committee reviewing a finished artefact.
Instrument the loop, not just the launch
Since outputs become future training data, monitoring has to cover the disparity metric over time, not only accuracy drift. Treat a widening gap the way you would treat a latency regression — as an incident with an owner.
Keep a decision record
Note what was chosen, what was rejected and why, at each of the five checkpoints. When a regulator, a customer or your own successor asks why the objective looks like that, a dated record is the difference between an answer and an argument. This is the part of responsible AI by design that overlaps with ordinary IT governance practice.
Where this connects to work already underway
Most teams meet this problem in the middle of a delivery programme rather than at a research desk. If you are shaping an AI strategy or reviewing how models reach production, the upstream checkpoints are the cheapest place to intervene, and the same discipline that protects a model from silent failure also protects it from silent unfairness — a point we made when covering why enterprises are rushing to secure AI deployments and again in the reporting on how enterprise AI actually learns in the field.
Responsible AI by Design Questions People Are Asking
Is responsible AI by design a new standard?
No. It is a methodological argument from the editors of a peer-reviewed journal, not a certification. ISO/IEC 42001 and the NIST AI Risk Management Framework are the nearest formal instruments, and the researchers’ point is that both work better when applied upstream.
Does adding fairness to the objective cost performance?
Not necessarily. In the health care study reported in the issue, the wait-time disparity was eliminated while scheduling efficiency was maintained. That is one system in one setting, so treat it as an existence proof rather than a general guarantee.
Should we just remove protected attributes from the model?
The scheduling result argues against it. Race had to be incorporated into the optimisation objective for the disparity to close. Removing the attribute typically removes your ability to measure a gap that correlated features will reproduce anyway.
Who owns this work in practice?
The five checkpoints Ramesh lists span data engineering, modelling, product and governance. Responsible AI by design fails when it is assigned entirely to the last of those four, which is the current default in most organisations.
Does the EU AI Act delay reduce the urgency?
It changes the deadline, not the design cost. Retrofitting an objective function after deployment means retraining and revalidation; doing it before the first training run means editing a specification.
How is responsible AI by design different from AI ethics?
AI ethics names the values. Responsible AI by design names the stage at which they have to be applied. The editors’ complaint is not that organisations lack values but that the values arrive after the objective function is written, when the only remaining remedies are the expensive ones.
Where does interpretability fit in?
It is the audit surface. If you cannot recover what a model optimised, and on what data, then nobody outside the build team can check any of the other claims. Interpretability is what makes responsible AI by design falsifiable rather than merely asserted, which is why the special issue treats it as one of the six domains rather than a nice-to-have.
What should a small team do first?
Compute one disparity metric for one system you already run, and write down the sentence describing what that system optimises. Those two artefacts will tell you more about your exposure than a policy document will.
References and Further Reading
Responsible AI must be built in from the start, researchers say — Tech Xplore
University at Buffalo School of Management news release
Responsible AI and Data Science for Social Good — INFORMS Journal on Computing
Ramaswamy Ramesh — University at Buffalo School of Management
Stanford HAI Artificial Intelligence Index Report
NIST AI Risk Management Framework
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.