AINL#021 Augmented Intelligence in Investment Management Newsletter
Welcome to the 021 Edition of the Newsletter on Augmented Intelligence in Investment Management.
Every two weeks, we deliver five unique insights tailored to empower investment decision-makers. Our insights are carefully curated by a seasoned team of market specialists. Unbiased, actionable and practical. They will help you navigate through noise.
AINL#021 SYNTHESIS
1. Gains in AI task automation will redefine productivity frontiers
The METR (2025) benchmark shows that AI task completion horizons have doubled every seven months for the past six years, implying autonomous completion of complex week-long tasks before 2030. This compounding trajectory signals significant opportunities for investors to back firms building domain-specific AI workflows, particularly in white-collar environments where automation can scale fastest (METR, 2025). Anticipating where these agents can be most effectively deployed—finance, research, compliance, will be critical to capturing the resulting productivity and cost-efficiency gains.
2. Trust, explainability, and risk management are becoming core investment filters
As Chen et al. (2025) demonstrate, interpretability methods such as sparse auto-encoders can make LLM forecasts increasingly traceable to economically meaningful features, reducing bias and enhancing regulatory readiness. Still not there yet: at the same time, Rubinson (2025) and OpenAI (2025) caution that AI hallucinations and misinformation risks are pervasive, with NewsGuard data showing chatbot falsehoods rising from 10% to nearly 60% in a year. For investors, this means prioritizing firms that embed explainability, bias mitigation, and uncertainty acknowledgment into their AI systems, both for compliance and to maintain decision quality in an environment increasingly exposed to unreliable automated outputs.
3. Climate and systemic risks demand integration into financial decision intelligence
The IMF (2025) emphasizes that climate change is a macro-critical policy challenge, with each 1°C rise cutting global GDP growth by 1.2 percentage points and disasters widening fiscal deficits by up to 10% in emerging markets. Combined with AI-driven decision infrastructures, this underlines a dual imperative for investors: backing innovations that enhance adaptive capacity to systemic risks (climate, regulatory, AI reliability) while integrating profit and impact motives at the portfolio level. Positioning capital where climate resilience and trustworthy AI intersect offers not only protection but also competitive advantage in volatile markets.
TOP 5 ARTICLES
ARTICLE ONE
Measuring AI Ability to Complete Long Tasks
ARTIFICIAL INTELLIGENCE | Metr | 8_2025 | Paper
Important Development
Multi-agent collaboration has emerged as a promising approach, enabling multiple AI agents to work together to solve tasks. This study by BlackRock researchers investigates the application of role-based multi-agent systems to support stock selection in equity research and portfolio management through a Fundamental Agent, a Sentiment Agent and a Valuation Agent. The system runs on Microsoft’s AutoGen framework using GPT-4o.
Why Relevant to You?
For investors, the implication is clear: while today’s benchmarks may not perfectly mirror real-world work environments, the acceleration suggests significant automation potential across white-collar domains.
In practice, this means opportunities to back firms building domain-specific environments and workflows optimized for AI agents, as well as to anticipate productivity gains and cost efficiencies where task automation can be scaled. The upside lies in identifying where these agents can be most effectively deployed—and positioning portfolios to benefit from the compounding trajectory of their capabilities.
ARTICLE TWO
A Financial Brain Scan of the LLM
HUMAN & ARTIFICIAL INTELLIGENCE | Chen et al. | 2025 | Paper
Important Findings
The paper proposes an approach to make large language models more interpretable. Using sparse auto-encoders and cluster analysis, the approach extracts the most relevant and economically interpretable feature groups that influence an LLM’s reasoning process. In the context of a financial use case the authors show that this approach can map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing. Furthermore, their results indicate that LLM forecasts are biased toward positive sentiment, but this and other biases can be addressed by the same mechanism that was used to extract the relevant feature groups.
Why Relevant to You?
The limited explainability of complex AI models poses significant challenges and issues for financial institutions and regulators. Explainability is essential for transparency, accountability, regulatory compliance and consumer trust. The same is true for the closely related concept of interpretability. The paper presents a simple, scalable approach to opening and controlling large language models. While this technique is still in its infancy, concept steering could evolve into a standard tool for building interpretable, field-ready LLMs for research and practice with applications across all social sciences.
ARTICLE THREE
Integrating Climate Change Into Macroeconomic Analysis
ARTIFICIAL INTELLIGENCE & SUSTAINABILITY | IMF | August 2025 | Paper
Important Findings
Different to the Worldbank, the IMF is now stepping up, integrating profit and impact perspectives, while utilizing latest quantitative finance techniques. Good to see. Climate change poses significant and diverse impacts on countries’ macroeconomic and financial stability, resulting in complex macro-critical policy challenges. This paper reviews (i) climate change and related policies’ channels of impact on the real, fiscal, external, monetary, and financial sectors over various time horizons and (ii) corresponding data sources, augmented models, and climate scenarios that could be applied in assessing the impact of physical climate risks as well as adaptation, transition, and mitigation policies.
Why Relevant to You?
- Carbon pricing or subsidy reforms may feel painful at first, yet they strengthen fiscal positions in the long run.
- A 1°C rise above historic norms can shave off 1.2 percentage points of global GDP growth.
- Disasters widen fiscal deficits by 5–10% of GDP in emerging economies within a few years.
Overall the paper is packed with insights and a good source for anyone trying to understand the concept and concluded implications.
ARTICLE FOUR
Chatbot Newsfeeds: 35% of falsehood and counting
HUMAN & ARTIFICIAL INTELLIGENCE | Rubinson S. | 09 2025 | Research Article
Important Findings
More and more people are looking for news through AI chatbots such as ChatGPT, Perplexity, Gemini or the likes. A new report by the American organization NewsGuard (that, among others, assesses news channels on their reliability) warns: in 1 in 3 cases, those chatbots spread false or misleading information. The research started a year ago and the researchers drew up a report every month. They find a growing trend in falsehood (from 10% to almost 60%). Selective ‘AI grooming’ seems to reinforce this trend even further.
Why Relevant to You?
A year ago, chatbots still seemed cautious. On sensitive topics, they sometimes refused to answer. Now they respond much faster and more extensively. In parallel, the number of mainstream chatbots has increased. Trust and use from people for private and professional newsfeeds increased as well. All factors combined systemically increase the risk of business decisions being taken based on false information. For now, Chatbot news should therefore be used with caution (implying eventually even corporate guidelines), and still in combination with traditional news sources.
ARTICLE FIVE
Understanding AI Model Hallucinations
Understanding why language models hallucinate | Open AI | 9_2025 | Article
Important Findings
A recent study from OpenAI examines the theoretical framework for understanding why large language models hallucinate. They analyze the pre-training origins, based both on quality of training data and statistical nature of the algorithms, and the persistence through post-training. The post-training analysis identifies evaluation misalignment as a core underlying reason for persistent hallucinations. Most benchmarks use binary scoring, rewarding guessing over expressing uncertainty, in much the same way that students optimize test scores by guessing under uncertainty when the score is simply based on right or wrong. This generates what the authors call an “epidemic” of penalizing abstention or uncertainty. The authors suggest that a way to mitigate this will require social and technical changes in evaluation design and the scoring of benchmarks. If the field purposefully shifts to evaluation designs geared toward rewarding expected behavior, including acknowledging uncertainty, it can lead to more trustworthy AI systems in the future.
Why Relevant to You?
Understanding hallucations as a normal statistical phenomena rather than mysterious failures helps frame appropriate use cases and safety measures for their use in investment decisions. For example, when conducting research, investment professionals should be particularly cautious about claims that seem plausible particularly when regarding rare facts, smaller or less-covered entities and recent events. The focus in this paper is on factual hallucinations, but it shows that even search and RAG-augmented models are not immune to these errors. Investment decisions involve multi-step complex reasoning and the framework presented applies only to parts of that process. More generally, adopting evaluation frameworks within teams that reward acknowledging uncertainty can align with more prudent risk management, particularly when working with LLMs. This paper gives a theoretical unpinning for the empirical results highlighted in the recent Newsguard study we highlight in Insight 4.
