At the Intersection of Ethics and AI

Here's the Problem in a Nutshell

• A 2021 Pew Research study found that a majority of experts and advocates worry AI will continue to focus on optimizing profits and social control and will not likely develop an ethical basis within the next decade. Not much has changed in the last five years. Admittedly, AI has made exponential progress at exponential rates since the study, but AIs still hallucinate, still make up sources, and routinely make horrendous recommendations,

• Fundamentally, AI Large Language Models (LLMs) have no real ethical foundation—whatever "ethics" AI LLMs display are the accident of their creators' predilections, biases, and specifics of education that gave them sufficient foresight to imbue them with by virtue of the sources chosen to populate a model or tuning done to exclude certain edge cases.

• AI-aided decision-making may work for you as a business or a government entity in the short term—until your customers and end users are skewered in the press by using your AI to determine benefits, support legal arguments, or make tactical warfighting decisions that create disastrous consequences.

• Even with model refinement and attempts to constrain consideration by the use of retrieval augmented generation (RAG), huge problems with AI will continue to relate to the RAG creators' and reviewers’ cultural biases, strengths, and motives, whether unconscious or determined by law, regulation, and policy—consider some real-life situations:

  • If a mandate is created that "No sources that advocate consideration of societal diversity can be included", or
  • A policy that "No psychological aspects associated with disability claims can be allowed to influence a determination", or
  • A military or immigration enforcement policy that advocates a “hunter-first” mindset: Identifying facial hue is a lot easier than identifying intent.

The bottom line is that we are already at a place where AI-assisted technology has done measurable and significant harm — deep fakes, revenge porn, credit theft, drone attacks on small boats have all been headline news. As the old, and possibly still-valid, adage goes, "No one wants to be a headline in the New York Times." The risks to reputation and the long-lasting trust deficit damages may be fatal to a business or an institution. Investors, consumers, and partners invariably react immediately, leading to public distrust, drops in stock value, or sales.

Where are the Fault Lines

• A lot of the problems in models relate to creators' and reviewers’ cultural biases, strengths, and motives.
• Remember that at the core, there is lingering doubt that AIs are little more than 'stochastic parrots': LLMs don't actually understand anything; instead, they are just massive statistical calculators that stitch together words based on probabilities—essentially repeating back human language without a shred of real comprehension.
• Simply hoovering up everything ever published bears the seeds of it's own destruction. Mein Kampf and The King James New Testament are given equal weight for inclusion in a LLM.

• Consider a thought experiment: suppose we had a weapons system that had a capability for judgment as good as that of a human (for generous values of as good as human )

  • What are the ethical considerations?

    • Who bears ultimate responsibility?

In short, just because you have a capability, you shouldn't operate without context. Ethics is about action guidance. AI must operate within a context. Also, the control/training/testing regime must closely model the situation likely to be faced in real life.

National Defence: AI and Moral Ambiguity

This is a domain with existential considerations. In the context of 2026, we are entertaining the concept of whether we might create a lethal autonomous weapons system (AKA LAWS), which is context independent. - Even in the hypothetical world, there would still be a context question - The value of taking a human life is (or should be) a weighty decision - If we’re taking ethics seriously, we should already have been talking about this for years. Bringing someone in at the last moment is a terrible idea. The fully autonomous self-driving vehicle quagmire clearly demonstrates the folly of bolting on an ethical context to a technical solution.

Where Does Moral Ambiguity Come Into Play

We must recognise that anything like a moral dimension exists in AI decisions and recommendations because of centuries of writing ingested into the model (and this can still be easily overwhelmed/overidden by ill-expressed prompts)

• An AI only concludes that an Iron Maiden is a bad thing because that opinion is supported by six centuries of writing and opinion. An AI cannot make the qualitative judgment that it's not the same as face recognition plus a trigger. We recognise the Nazi gas chamber as an evil technology in and of itself because the available literature says it is. A lot of lessons from drone warfare do not have such value assessments attached because 1). The assessment appears to depend on goals, and 2). There is moral ambiguity attached.

• When it comes to the moral dimension, several questions attach:

  • Can X operate in a real-world environment in exactly the way we intend it? No plan survives first contact — no AI is that smart
  • What did humans use to do based on intuition and judgement that now we need AI generated rules for? What is the quantitative definition of “terrorist”. What happens when someone approaches a checkpoint seems to be unarmed but won’t stop
  • Are there some decisions we never want machines to make. DoD doctrine has held since early days that only humans are legal agents.
  • What about the opinion that there are decisions we "never want machines to make.” A couple reasons — no right answer. No clear metric for “better” as in specturm allocation. What kind of world where “no one is responsible” does it degrade morality?
  • is there a moral difference between Hiroshima vice letting a drone linger and make a decision autonomously
  • The more remote from the target, the less responsibility one may feel. Consider a drone pilot sitting in Vegas, shooting in Iran. Patriot drone pilots believed that if they overruled the machine they would be disciplined. It reinforced bias

• If we are comfortable with the actions, then there’s still the situation where we still want to have a thoughtful culture in place. Responsibility needs to stay in the chain of command. Else war becomes too easy.

  • Carpet bomb: certainty is too difficult to achieve;
    • Loiter drones: who knows what happens between the “go” decision and payload delivery? Are we comfortable with drones that can linger forever and hunt down people?

• In a converse world where we have agentic AI smart machines, and must consider sending out humans to do the killing. Is one paradigm more likely to be better than the other? Some ideas that we should consider (which are not incorporated in any AI, agentic or otherwise):

  • It’s always important to have “skin in the game." But what happens when there is a non-human in the decision loop
  • War is terrible, but we have a clear historical read on when and how people can and will fail.
    • We have a theory of mind for humans, but most emphatically NOT for machines. Machines do not 'think like us'. In fact, they do not 'think' at all, at least in ways that can accurately be predicted, even when we see Extended Thinking mode in our conversational bots.

Autonomous Vehicles, Consequences of Ethics-free Real-time Decisons

Returning to fully autonomous self-driving vehicle for a moment, consider that if they could save 90% of lives, should that be the end of the story? Consequentialists would say yes. But others might ask, "Who is included in the 90%? Who is excluded?" Here's the non-trivial problem: we don’t yet know how they might fail and for whom they might fail.

The Well-known 'Colour' Problem

As is known , computer vision is badly trained against people of colour. Now imagine we slash deaths by 90% but every death is a small black child. Is it still worth it? Or is the answer to train better to create a better outcome?

Agentic AI in Health and Welfare Decision-making

Interestingly, much of what was pointed out above regarding national defence and the conduct of war applies to health and welfare decisions as well. We are already seeing the results of the rise of AI in benefits administration—particularly in health insurance and Medicare—this has led to a surge in automated, rapid claims and prior-authorization denials. While intended to increase efficiency, these systems have drawn massive scrutiny for overriding human judgment and causing dangerous care delays. The Impact of AI Denials is seen across several dimensions:

• Speed and Scale: AI algorithms can process and deny claims in as little as 0 to 17 minutes, rejecting batches of hundreds at a time.

• Patient Harm: Physicians report that AI-driven prior authorizations frequently deny medically necessary treatments, forcing doctors to waste hours on endless appeals instead of seeing patients.

• Low Appeal Rates: While roughly half of all denied claims are successfully overturned on appeal, less than 1% of patients and providers can actually navigate the complex appeals process.

The primary criticisms and rationale for this state of affairs is very similar to those cited above:

• Lack of Context: Algorithms often evaluate claims using rigid guidelines without considering a patient's unique, individualized medical circumstances. Ethical AI will consider context and the human dimension.

• Historical Bias: Because AI models are trained on past insurer data, they risk industrializing and locking in historical biases. This occurs even in RAG constrained decisions.

• Corporate Incentives: Critics, including major lawmakers and the American Medical Association, argue that AI allows profit-driven insurers to cut costs by effectively rubber-stamping blanket denials.

Human Health and Welfare will continue to be, in the U.S. at least, a major political controversy, and as agentic AI offers the tantalising prospeft of removing humans from the decision making process all over the decision chain, the situation becomes even more crucial to consider.

What Else Is Hard About AI and Ethics?

First, we should remember there are no perfect data sets whether in war or peace where everything went right. Maybe humans don’t do things that badly; maybe we just don’t have enough data to assert this.

We still don’t know how to create a machine that weighs moral values, and what those moral value ought to be. We have different value judgments as humans and as societies that aggregate humans.

One thing we know about ourselves as a species—human have the ability to muddle through. Agentic AI cannot claim the same latitude for their judgements. But remember the lesson of the Milgram Experiments: humans can be manipulated into acting in ways that they might not otherwise consider moral.

When it comes to things that affect the application of laws, we do have general and universal norms. However, the just outcome is deeply involved in the argumentation between individuals. What’s important in this type of system is that I am not automating my peers. There is no value-free way to make the machine a decision maker.

And of course:

• We miss stuff as software developers—and developers (and now increasingly developers using AI) are the mechanism to create agentic AI systems. Training and education in the technical disciplines must increasingly incorporate such 'soft' disciplines as philosophy and literature, and organisations must consider employees' and candidates' prior training or experience in these areas as critical success factors
• Autonomous systems like vehicles: would you rather be killed by a drunk driver vice a self-driving car. With a machine, we might never know a reason for the failure — but does that matter?
• Imagine, though, 8,000 years ago if we were deciding to domesticate animals, woud we have made the same decisions asto-day?

The Guiding Philosophy of the ECE-CAIR ( Electrical and Computer Engineering-Center for AI Research) at CUA

We are well beyond the realm of AIs being laughed off as "stochastic parrots". The analogy fails because we are in the realm of emergent capabilities as we cross over from decision support to decision implementation.

The reason the "stochastic parrot" argument has lost some traction is that as AI models have grown larger and able to absorb more tokens, they are beginning to demonstrate capabilities that simple statistical mimicking cannot explain. Here is some of what we are experimenting with at CUA's ECE-CAIR.

1. Internal Functional "World Models" To predict the next word accurately in a complex text, an AI can't just look at word frequencies; it has to understand the relationship between the concepts behind those words. Researchers have found that AI models build internal "world models." For example, if you train an AI on text descriptions of board games (like Othello or Chess), the AI internally builds a dynamic, 2D map of the game board to predict the next legal move, even though it was never given a visual grid or explicit rules. It infers the underlying structure of reality from the language.

2. Novel Problem-Solving and Logic If an AI were just a stochastic parrot, it could only repeat or slightly remix things it has already seen. However, modern models can solve completely novel logic puzzles, write entirely unique software code to solve specific tasks, and apply reasoning frameworks to situations that do not exist anywhere on the internet.

3. Cross-Domain Synthesis A stochastic parrot cannot read a medical research paper in English, synthesize the data, and translate it into a Shakespearean sonnet written in Spanish. An AI can. This requires an underlying, abstract representation of meaning (semantic space) that allows the model to map concepts across completely different domains, languages, and modalities.

Even as we have made significant gains in very short order, the current state of the art suggests significant future work, which the lab is well equipped to perform at reasonable costs. Considering where we are to-day helps our ECE-CAIR multidisciplpinary labs set a course for to-morrow:

Current State of the Art

Current State of the Art

Feature	The Stochastic Parrot	Modern AI System
Mechanism	Matches words to nearby words based on surface-level statistics.	Uses deep attention mechanisms to understand context and abstract concepts.
Creativity	Shuffles existing human phrases.	Generates novel code, poetry, and logic solutions to unique prompts.
Understanding	Zero concept of the underlying reality.	Possesses functional "world models" (e.g., understanding tracking, logic, and spatial rules).
Limitation	Stochastic Parrots can't reason.	Excels at functional reasoning, but struggles with abstract human consciousness and objective truth, and in particular fails to embody what humans commonly understand as 'ethical judgment'

In Conclusion

Perhaps having to face the problems with machines forces us to face it with humans, and that is a part of our institutional fibre, our instituational charter, our moral makeup. History has taught us that we can handle human “messyness” but what about agency in machines? Consider children — little autonomous agents any of whom can grow to be the next Hitler or the next Mother Theresa. We believe we know how to socialise, and school our kids, but that takes many years and a moral foundation—years which the market may not have or allow itself to have. Also humans are predictable in large numbers. Statistics at the individual level are far less predictable and impossible for us to add absolute guard rails. Agentic AI is unpredictable, both in the aggregate and in repeated trials using that same challenge questions on successive days.

Can we claim we are even close to respecting and embedding genuine ethical choices. As we have suggested above, they are made by one of a very small number of providers, for global products, without reference to any single organisation or society that will later run it. At ECE-CAIR , we see that the answer is currently a resounding no.

In humans, we can claim predictability and claim we can achieve some level of 'fairness'. We claim we can mitigate racial profiling, or social pigeonholing, but it's an act of will, and if nothing else, the current political climate here and abroad suggests that acts of will can change quickly and dramatically. We are an institution that, unqiue in academia, considers humanity to be our mission—both in physical domains and cyber domains.