Nonprofit AI Considerations

Ethics

There are important components to consider before any nonprofit implements AI in its fundraising strategies. For example, if winners write history and Large-Language Models (LLMs) are trained on historical data, who tells the stories of minoritized groups? The internet reflects a collective “us,” with anonymity often encouraging people to show their bad side. The LLMs are trained on vast amounts of data from websites such as Reddit and Wikipedia, and while most websites do not directly encourage bigotry, hatred, misogyny, racism, or any other bias, the overall tone and explicit language on the internet form the majority of the training data for these models. While the creators take precautions to avoid harm or spread bias, the imperfect models can generate offensive and non-inclusive text and/or images.

Since the training models are black boxes, managing ethical risks and bias is more complex with Gen AI. In their article, “Ethics and AI: 3 Conversations Companies Need to Have,” Reid Blackman and Beena Ammanath recommend understanding the sources of the AI models to create better strategies to mitigate bias.

“Understanding bias in AI requires, for instance, talking about the various sources of discriminatory outputs. That can be the result of the training data; but how, exactly, those data sets can be biased is important, if for no other reason than that how they are biased informs how you determine the optimal bias-mitigation strategy.”

Privacy

The data used for predictive models or content generation can be used by companies offering prediction platforms to further train their models, a significant problem if sensitive data from other parties is used on these platforms without consent. There are guardrails to protect sensitive information, such as using trustworthy providers and declining to share your data with the provider for further training. Another concern is the ownership of the data used for training, as copyrighted material such as news articles or stock images forms a good chunk of these training datasets. Some companies have sued OpenAI and other generative AI companies for breach of intellectual property.

Accuracy

As the famous saying goes, “All models are wrong, but some are useful.” Mathematical models try to approximate a customer’s following action but will never be 100% accurate. Similarly, the LLMs can not only get the math wrong, they can also get the facts wrong. “Hallucination” is an industry term for this problem where an LLM confidently gives incorrect information. It is a misnomer: these models simply make things up, and the term hallucination minimizes the risks. If the LLM-generated text must be fact-based, then every word must be verified. As one lawyer found out, one can risk their prestige and profession by unthinkingly trusting these models. The Federal Trade Commission is also investigating the harms of these inaccuracies.

While Gen AI can quickly analyze large amounts of information for new insights, it can spread bias and misinformation. Authors Henry A. Kissinger, Eric Schmidt, and Daniel Huttenlocher caution in their book The Age of AI: “AI’s effects on human knowledge are paradoxical … AI is capable of exploiting human passions more effectively than traditional propaganda. Having tailored itself to individual preferences and instincts, AI elicits responses its creator or user desires. Similarly, the deployment of AI intermediaries may also amplify inherent biases, even if these AI intermediaries are technically under human control. The dynamics of market competition prompt social media platforms and search engines to present information that users find most compelling. As a result, information that users are believed to want to see is prioritized, distorting a representative picture of reality.”

Do you think the benefits of AI in processing and analyzing data outweigh the risks of misinformation and bias?

Yes, the advantages surpass the risks.
No, the potential for harm is too great.

Talent

Automation has reduced many repetitive, often mundane, tasks. These new AI-powered generators promise to reduce many jobs’ workloads. One OpenAI and University of Pennsylvania working research paper suggests that “80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs.” While every innovation has increased worker productivity and created new tasks for humans, AI can feel different; creativity has been the domain of humans, but AI has now breached it. If a job involves repetitive tasks requiring reading and responding to text without much critical thinking, it seems like it will be the first to be augmented via LLMs. In addition, creative tasks to generate evocative and original images seem to be under threat of replacement, as “A machine can’t do this” seems like a phrase from the past. However, it is important to remember that people are often willing to pay a premium for an expert’s specialized knowledge.