AI Bias In Investing Cannot Be Eliminated, But It Can Be Governed

Whilst AI is fast becoming the norm for many investment firms, its increasing adoption is not without its problems. Current Large language models (LLMs) can summarise research, interrogate datasets, draft investment committee papers and support client communications. These gains in speed and reach are significant and in general beneficial but if not monitored closely such speed can also allow and facilitate the flawed reasoning to travel further and faster.

Although it is tempting to regard an LLM as neutral, every version reflects choices made by humans, teams, and organizations embedded in its training data, architecture and instructions. The investment workflow then adds another layer of human judgement: how a question or prompt is framed, which information is included or omitted, which tools are used and whether the output is challenged. The risk is therefore not just the model itself, but the entire human / AI system and the interaction between them.

Framing is a particularly important example as identical financial information presented in terms of gains rather than losses can lead an LLM to reach a different assessment. In investment settings, this effect can be especially pronounced when discussing drawdown performance or client relationships, where the language used and method of communication can carry emotional and behavioural weight.

READ | 15 Million Jobs At Risk From AI? Goldman Sachs Has A Timeline Involving The Recession

This matters because investment decisions are consequential. A small alteration introduced at the research stage may influence the final output such as a scorecard, committee paper, portfolio recommendation and subsequent client conversation. If one model is utilised by several analysts, the same bias can be repeated at scale unchecked and even unnoticed.

Curiously defining a bias or instructing an LLM to avoid it generally has little effect as even though, in principle, models can recognise bias they can still reproduce it, in practice. Hence firms should look to embed testing, workflow design and human accountability in their processes.

Ideally, the practitioners and decision makers should look to test workflows rather than individual outputs; comparing responses to paired prompts that contain the same facts but varying the framing, order, labels or context. Repeating those tests across different scenarios and models allows you to test if, and where, conclusions change and the reasoning involved.

This will not only help in understanding how the model is working, but also in clearly articulating what needs to change with a better alignment with end objective and correction of ingrained principles, when needed.

Standardising prompts for frequent tasks is a good practice and as such, a templated approach allows you to observe where the model makes assumptions or inferences, which can then be flagged and, where recurring, and can be adjusted going forward. Such an approach will not necessarily remove bias, but it will make the reasoning easier to inspect and challenge.

Given the current limitations of LLMs, human oversight remains essential, particularly for decisions with material portfolio, client or conduct consequences. Yet "human in the loop" must mean more than approving a model's answer and reviewers should first understand the end-to-end process so that they can restructure prompts, challenge source selection, test alternative interpretations and override the system when necessary.

READ | October Tech Layoff Wave: Over 5,000 Jobs Gone As Amazon, Netflix, eBay, HubSpot Cut Staff

It's also important to acknowledge that humans have biases too and diversity of opinion in taking these judgments may be helpful; the purpose is to introduce informed challenge and accountable judgement, not to presume human infallibility. For high-impact and frequently used applications, firms should consider developing internal bias benchmarks that allow new models and major updates to be tested before deployment.

Governance should also explore and distinguish between implicit biases, which arise from pre-training and model design, and explicit biases revealed through observable choices such as data selection, tool use or a proposed action. Reasoning traces and audit trails may illuminate the second category, but only close scrutiny and testing of the output will reveal the first.

All these additional skills need to be developed to a point where investment professionals' AI literacy extends beyond prompt writing to aspects including behavioural finance, model risk, experimental testing, data provenance and professional scepticism. As early career roles adapt to an AI-enabled world, the most recent entrants such as junior analysts need to be encouraged by more senior decision-makers to question authoritative-sounding output.

Bias cannot be fully eliminated from either humans or machines but there are ways to make it detectable, measurable and manageable. Firms should ask where bias can enter, how it will be found, who must challenge it and what evidence is required before acting. These practices apply wherever investment firms use AI, but how far firms must go is increasingly shaped by local regulators, and India offers a useful example.

The regulator's role is also very important. India's capital markets regulator, SEBI and its central bank, RBI, maintain close attention to AI deployments. The RBI's FREE-AI framework sets out recommendations for AI governance in the financial sector, while SEBI's June 2025 consultation paper proposes guidelines for capital markets. Common principles include:

● Governance: Boards and senior management are expected to play an active role in AI oversight, including through board approved AI policies

● Privacy and Data: AI's reliance on data ties it closely to privacy and residency regulations.

● Explainability and Fairness: Transparent, non-discriminatory outcomes are framed as central to trust

● Vendor Accountability: Firms remain accountable for AI they source from vendors, and FREE-AI recommends AI specific safeguards in outsourcing contracts, including on bias.

●High-Risk Applications: In credit decisions and insurance underwriting, supervisors are alert to both the promise of efficiency and the model risks, training bias, monitoring.

The good news is that investment profession and the broader industry have long understood that trust rests not only on outcomes but also on process, and this same principle should govern AI adoption. Those that treat bias as a purely technical defect will miss the point while those that build transparent, testable human-and-AI workflows can capture the technology's advantages without surrendering judgement.

This story is authored by Brian Pisaneschi, CFA Director, Applied Investment Practice & Tools CFA Institute

Disclaimer: The views expressed in this article are solely those of the author and do not necessarily reflect the opinion of NDTV Profit or its affiliates. Readers are advised to conduct their own research or consult a qualified professional before making any investment or business decisions. NDTV Profit does not guarantee the accuracy, completeness, or reliability of the information presented in this article.