AlphaFold and Insilico: AI Redefines Science

Collaborative AI Systems: Human-AI Teamwork Workflows

In the world of innovation, some pioneering companies have adopted a collaborative approach between artificial intelligence and humans to make informed decisions. AI generates options and highlights patterns, while humans provide their judgment and context to finalize decisions. This model of collaboration does not just give orders but involves continuous interaction between both parties.

Observing Real-World Applications

This collaborative model is not just theoretical; it is already in practice across various sectors.

Transforming Scientific Research and Healthcare

AlphaFold, for example, has revolutionized research in molecular biology by predicting protein structures in a matter of hours, a process that would have taken years in the lab. However, the interpretation of these predictions and the planning of necessary experiments remain the responsibility of scientists. The biotechnology company Insilico Medicine has pushed this approach even further. Traditional drug development takes four to five years just to identify a promising compound. Insilico Medicine has built an AI platform that generates and filters thousands of potential drug molecules, predicting which ones are most likely to succeed. Then, medicinal chemists review the best candidates, refine the structure, and create experiments to validate them. The results have been significant: the time required to discover a lead compound has decreased by about 75% — from four or five years to just 18 months.

The same pattern exists in pathology. PathAI analyzes tissue samples to diagnose diseases like cancer. Pathologists then review the AI's results and add their clinical experience to establish a diagnosis. According to a study from Beth Israel Deaconess Medical Center, the outcome was 99.5% accurate cancer detections, compared to 96% when the pathologist examined the slides independently. Moreover, the time required to review the slides has significantly decreased. AI detects patterns that fatigue might have caused to be missed; humans provide the clinical context.

What we have learned is that AI finds patterns — it excels in volume and speed. Humans excel in judgment and context; they determine whether these patterns are significant.

Enhancing Business Decisions

AI can accomplish in a few hours what used to take teams weeks: reviewing thousands of contracts, analyzing risks across global markets, and identifying patterns in usage data. All of this can be done quickly, but deciding what to do with this information remains a human responsibility.

For example, legal teams at JPMorgan Chase manually reviewed contracts for 360,000 hours each year, a slow, costly, and error-prone process. They created a solution called COiN, an artificial intelligence platform designed to read legal documents through natural language processing (NLP) and machine learning. COiN can extract key points from legal documents, identify unusual or questionable clauses, and categorize provisions in seconds. However, lawyers still review the items flagged by the system. As a result, JPMorgan can process contracts much faster than before, reduce compliance errors by 80%, and allow its lawyers to spend their time negotiating and developing strategies rather than repeatedly reading contracts.

In another example, BlackRock is the largest asset manager in the world, controlling assets worth a total of $21.6 trillion for institutional clients and individual investors. At this scale, BlackRock must analyze millions of risk scenarios across multiple global markets, which cannot be done manually. To address this issue, BlackRock developed Aladdin (Asset, Liability, Debt, and Derivatives Investment Network), an AI-based platform to collect and process large amounts of market data and identify potential risks before they occur. There is always a human component: BlackRock's portfolio managers review Aladdin's analyses and make all allocations. The results show that risk analysis that previously took days is now done in real-time. Moreover, BlackRock portfolios created using Aladdin's analyses, combined with human judgment, have outperformed both purely algorithmic and purely human approaches. Currently, over 200 financial institutions license the Aladdin platform for their own operations.

The pattern is clear: AI sheds light on options and information at scale. But it won't tell you when you're wrong; you'll have to figure that out yourself. JPMorgan's lawyers still review what COiN flags, and BlackRock's portfolio managers still make the final decisions.

Examining Collaborative AI Tools

Not all AI tools are designed for collaboration. Some provide output in a "black box" format, while others have been created to collaborate with you. The list below highlights tools that support collaboration:

Using General-Purpose Assistants

• Claude / ChatGPT: These are conversational AIs that provide feedback on your reasoning, point out ambiguities, and will tell you when they are unsure. They represent the tools closest to true interactive collaboration.

Conducting Research and Analysis

• Elicit: This tool searches academic articles and extracts results, showing you the evidence behind claims so you can determine whether to accept the information.

• Consensus: This platform synthesizes scientific literature and displays areas of agreement and disagreement among researchers so you can see all aspects of a discussion.

• Perplexity: This provides search results with citations. Each claim is linked to a verified source.

Optimizing Coding and Development

• GitHub Copilot: This tool suggests code completions. You review, accept, or modify; nothing executes without your approval.

• Cursor: This is an AI-native code editor. It displays the differences in proposed changes so you can see exactly what the AI wants to modify before it happens.

• Replit: This provides explanations for code, suggests corrections, and assists with debugging. You remain in control of what is deployed.

Advancing Data Science Workflows

• Julius: This tool analyzes data and creates visualizations. It displays the code used to create the visualization so you can audit the methodology.

• Hex: This is a collaborative data workspace with AI assistance. It was created for teams where humans and AI work together on analysis.

• DataRobot: This is an automated machine learning (AutoML) platform that provides explanations for model decisions. It displays feature importance and prediction confidence so you understand the underlying logic.

Enhancing Writing and Communication

• Notion AI: This tool is integrated into your workspace for drafts, summaries, and brainstorming sessions, but you choose what stays.

• Grammarly: This provides modification suggestions with explanations. You accept or reject each individual change.

What makes these tools collaborative is that they show their work. They allow you to verify their results and do not require you to accept their output. This is the difference between a tool and a collaborator.

Measuring Collaborative Success

Three types of metrics help you evaluate whether human-AI collaboration is truly working:

• Outcome metrics: Easy to track. Are you getting better results? Faster turnaround times? Fewer errors? You should monitor these indicators.

• Process metrics: Even more significant. If you are never rejecting AI outputs, that is not a sign of high-quality AI; it is a sign that you have stopped thinking.

• Human experience: Is it possible to produce these results without AI? Do you really understand why the AI made its choices, or are you just following along because it seems smart?

A good check: if you always accept the first output, it feels more like stamping than collaboration. Working without AI from time to time helps you maintain a baseline so you know what is your work and what is the tool's.

Implementing Effective Practices

Teams that successfully collaborate well with AI tend to follow a few common practices:

• Establish clear roles: Determine what role you play and what role the AI plays. A common setup involves the AI generating options while you select the best one. This allows you to leverage AI's ability to explore many possibilities while keeping the final decision with you.

• Integrate checkpoints: Do not let AI outputs move directly to the next phase without a brief pause. You do not need formal approval, but you should take a minute to reflect on why the AI made its choice. If you cannot articulate the reason, do not accept the output.

• Demand transparency: Use tools that show their work, including the code they generated, the sources they used, and the modifications they proposed. If you cannot see how the AI arrived at its output, you cannot verify it.

• Stay sharp: Work periodically without AI. This is not a statement of resistance but rather a standard for comparison. You want to know what your work looks like without assistance, and you want to be able to perform if the tools fail.

Concluding Thoughts

Human-AI teamwork represents a true shift. We are learning to interact with systems that provide input rather than simply executing commands.

For this to work, new skills must be acquired, such as knowing when to trust AI and when to challenge it. This involves evaluating processes to see if they produce results or merely seem productive. Most importantly, it requires staying vigilant enough to spot errors when they occur.

Teams that develop ways to collaborate with AI produce better outcomes. They identify errors earlier and consider options they might not have otherwise contemplated.