Ideas

This post explains what people often mean when they say “AI” and why the differences matter for foundations, nonprofits, and evaluators. It compares predictive machine learning, generative AI, and precision causal modeling, showing how each tool supports different decisions, requires human judgment, and carries different risks.

If you read Kien Lee’s post on evaluating and learning faster with AI-powered analytics, or joined our webinar on the same topic , you may have noticed that we used the term “AI” often without stopping to define it. That’s understandable: AI has become a catchall for many different technologies, from recommendation engines to ChatGPT and other tools that can summarize information or support analysis. But when one term is used to describe so many different things, it can obscure the practical differences that matter most: what the tool is designed to do, what kind of judgment it supports, and what risks come with using it.

A Short History Helps Explain Today’s Confusion

Part of why AI feels so slippery is that it has meant different things at different times. Researchers often describe the field in four waves.

The first wave, roughly the 1950s through the 1980s, was the symbolic era: AI as rules written by humans, logical “if-then” statements that a machine executed. This approach was transparent and explainable, but it did not scale very well to messy, real-world data.

The second wave, the 1990s into the early 2010s, shifted from coding rules to learning patterns from structured data. This is statistical machine learning: spam filters, recommender systems, and models like random forests that could handle uncertainty better than rigid rules, but were still narrow and built for one task at a time.

The third wave, around 2012 to 2017, was deep learning. Neural networks finally worked at scale, thanks to more powerful computer chips and much larger datasets. This is when computers started to “see” and “hear” with human-level accuracy, powering advances like speech recognition and game-playing systems that could beat human champions.

The fourth wave, from 2017 to today, is generative AI. A technical breakthrough called the transformer made it possible to build foundation models — systems trained on enormous amounts of text, images, or other data that can synthesize, create, and reason across many different tasks. This is the wave that gave us ChatGPT, Claude, and similar tools, and it marks a real shift from AI that classifies information into categories to AI that creates new content and reasons through problems.

Knowing this history matters because it explains why AI can refer to something as simple as a decades-old spam filter and something as complex as a modern chatbot. They are built differently, they behave differently, and they carry different risks.

Three Forms of AI-powered Analytics You’re Likely to Encounter

For organizations doing evaluation, case management, or strategy work, three forms of AI-powered analytics are especially relevant right now.

Predictive machine learning uses historical data to identify patterns and forecast future events. Its central question is: Who is at risk? Imagine you have data on a person: past shelter stays, service use, recent moves, health needs, income history, and current living situation. A predictive model looks across thousands of past cases to find combinations of factors that showed up before longer homelessness, a return to crisis, or higher service needs. It then compares your case to those patterns, not to judge, but to estimate what is more or less likely to happen next. The important caveat: a prediction is not a destiny. It tells you someone matches patterns associated with a certain outcome; it does not determine that outcome.

Generative AI creates new content and synthesizes complex information. Its central question is: How do we communicate and act? Here is a simple example. You give a large language model a case note, an agency policy, and a question about the case. The model has been trained on huge amounts of text by repeatedly predicting the next word in a sequence, and using that training, it produces a written response sensitive to your note and policy. That output might help draft an email or contribute to an action plan. The caveat here is critical: these models can be wrong, sometimes confidently so, and therefore human review is always necessary before you act on their output.

Precision causal modeling moves beyond correlation to ask about cause and effect. Its central question is: What works, for whom, and under what conditions? Instead of just predicting an outcome or generating text, this approach starts with a target, such as sustained stable housing, exiting foster care, or achieving some form of financial well-being, and finds people with genuinely comparable starting conditions and context. It then looks within that group to identify which services or supports were actually linked to better outcomes for people like them, surfacing what might work for a specific subgroup rather than a generic, one-size-fits-all recommendation.

These three forms answer different questions, and that distinction matters more than the umbrella term AI ever will. A tool built to predict risk is not the same as a tool built to draft text, and neither is the same as one built to identify what actually causes better outcomes for a specific group of people.

Why the Distinction Matters for Your Evaluation, Learning, and Strategy Adaptation

Kien’s post made a version of this same point: the more useful questions are still: What are we trying to learn, and which data will help us answer that? That question is impossible to answer well if you don’t know which kind of AI tool you’re talking about.

If your organization needs to demonstrate impact or understand why outcomes vary across providers or populations, you’re likely looking at causal modeling, not a generic predictive score. If you’re trying to help staff draft communications or synthesize case notes faster, that is a generative AI use case, and it comes with a specific obligation for human review because the technology is fundamentally producing plausible-sounding text rather than verified fact. If you’re trying to flag who might need more intensive support soon, that is predictive analytics, and it should be treated as a starting point for a conversation, not a final determination about a person’s future.

Conflating these categories is where a lot of AI adoption goes wrong. An organization might apply generative AI’s speed and fluency to a task that actually calls for the accountability of causal modeling, or treat a predictive score as if it were a settled fact about a person rather than a pattern match worth investigating further.

A Few Ground Rules for Responsible AI Use

Whichever form of AI-powered tool you’re considering, a few principles apply across the board, and they echo what panelists described in the webinar.

Human judgment stays essential at every stage. Someone has to define the criteria used to organize data, test the AI’s output against a manually reviewed sample, and review the final analysis for accuracy and context. The technology accelerates the work; it does not replace the expertise behind it.

Start with the question, not the tool. Decide what you are trying to learn and which data can actually answer that question before you decide whether an AI-powered approach is worth the investment. More data and more processing power do not automatically produce better decisions.

Expect upfront investment before efficiency. Systems, staff training, and validation processes all take real time and resources to set up. The payoff tends to show up later, in an organization’s capacity to learn and adapt continuously, not as an immediate cost reduction.

Validate, don’t just trust. Predictions can be wrong. Generated text can be wrong. Even causal findings depend on how comparison groups were constructed. Build in a step where a person checks the output against both the data and lived expertise before anyone acts on it.

The Bottom Line

AI is not one thing. It is a history of approaches stretching back 70 years, and today it shows up in our work mainly through three distinct forms: predictive analytics that estimate risk, generative AI that drafts and synthesizes content, and causal modeling that identifies what actually works for specific groups of people. Knowing which one you’re using, and why, is the first real step toward using any of them responsibly.

As Kien’s post argues, the goal is to strengthen human capacity to ask better questions, validate what the technology gives back, and make better decisions with the people the work is meant to serve. Understanding what these tools actually are, and are not, is where that starts.

About the Author

• What does AI mean in evaluation, strategy, and nonprofit work?
• Why is AI such a broad and sometimes confusing term?
• What is the difference between predictive machine learning, generative AI, and causal modeling?
• How can organizations decide which AI-powered tool fits their question?
• Why should predictive analytics be treated as a starting point rather than a final decision?
• Why does generative AI require human review?
• How can causal modeling help organizations understand what works, for whom, and under what conditions?
• What ground rules should organizations follow before using AI-powered tools?