I had the privilege of sharing the stage with a lovely group at the 2023 FIA Conference to discuss different perspectives on Machine Learning.
I heard after the session that the room was so full that the doors were shut and people couldn’t get into the session. It’s great to hear that it’s a topic that many fundraisers are interested in and I wanted to share my responses (as best I can recollect after a big day) for those that may have missed out.
Have you used Chat GPT before in the Fundraising context? Can you tell us about it?
We provided a brief example of a letter that we asked Chat GPT to write from Vinnes QLD asking me to provide support to help end the housing crisis in Queensland.
The response was a basic letter that had some positive aspects:
- noting Vinnies’ more than 100 years of service to build credibility
- the use of emotive language to demonstrate the need
- Letting me know that I can make an impact
It wasn’t a great letter, but it was a reasonable foundation for the minute it took to generate it.
Because Chat GPT is a conversational interface, I asked it to update the letter to include a Variable Paragraph and then revise it for a different audience. We could also ask it to generate email subjects and copy for social ads by asking it for a little more.
This sounds like a dream for time-poor fundraisers, and if something sounds too good to be true, it usually is. So what’s the catch? It comes down to the unknowns. We don’t know if the email subject lines will pass our Mail Chimp spam detection or not. This extends to a potential future problem, where we observe an arms race between AI content generation and Spam detectors, where each is trying to outdo the other to deliver for its core users. As fundraisers, we’re probably not prepared for our message to be filtered out from donor inboxes because we used some AI to generate our content.
Is AI/ML a strategic tool or a short-term tactical tool?
In its current application, AI/ML is demonstrating itself to be a primarily tactical tool. This is because the core premise of AI/ML is pattern matching. The model learns from patterns observed previously and makes suggestions based on those inputs.
If we take the example of some Major Donors at Tax time, where we’ve received some very kind and generous gifts from supporters, we might be inclined not to include a Response Mechanism in the next impact newsletter. This results in a pattern in our data that shows a lack of giving from High-Value supporters in the Spring period. Not because they weren’t willing, but because they weren’t provided with an opportunity to respond. When we run our model next spring it identifies that our major donors aren’t making donations then, and suggests that we don’t need to target them. As fundraisers, we’d probably agree that this isn’t a wise donor stewardship approach for hour High-Value supporters.
At this point, while AI/ML is only able to identify and replicate patterns, it won’t be able to create anything ’new’ or ‘unique’ which is what’s needed in strategic applications.
ML is very good at answering specific questions, is it used to answer the correct questions for fundraising?
At the moment, the models I’ve observed in the AI/ML space focus on short-term questions like, will this donor churn in the next 3 months, will this donor give in the next 3 months, will this donor make a high-value gift in the next 3 months? These are short-term questions that address short-term needs. They can however have negative long-term side effects. A typical question of will this person respond in the next 3 months, doesn’t take into consideration how much they’ll respond with, and whether they’ll make a positive return. It also may exclude some new donors because they’ve only made one gift. It doesn’t take into account the fact that you’ve invested substantial sums to acquire those donors, and aren’t allowing them to respond. Based on the concept of the Rule of Seven our new supporters aren’t getting all the available opportunities to activate. This can lead to many of the factors we’ve known are important (second gift rates & retention) in Fundraising starting to decay. It also means that next time round when we ask ‘will they respond in the next 3 months’, they’re probably not going to be identified with a good propensity, because their recency is a bit less. By asking a short-term question, we’ve had a slight improvement in our efficiency, but have limited our lifetime value opportunity.
The ultimate question we should try to be answering is, “Should I send this supporter this appeal to maximise their lifetime value”? The problem is, this is a much harder question to answer because of the volume of data required, the longevity of data we need to have to see enough communications, and the quality of data that would need to be maintained to answer the questions.
What should fundraisers watch out for when thinking about using AI/ML?
There are several key factors to consider.
- Are there any longer-term side effects that we can anticipate before we implement AI/ML models at scale?
- Are there any biases that will be perpetuated through re-training? One example of this is the UK Home office implemented an AI algorithm intended to streamline immigration visas. Unfortunately, some early inputs about alleged ‘high-risk’ countries ended up having small biases increasing over time with re-training. You can find out more about this here
- On a more ethical side of things, we should be considering whether the use of AI/ML models is an appropriate fit for our purpose. Environmental charities may wish to weigh up the environmental costs of using tools like Chat GPT. It’s been estimated that to do one iteration of training of GPT3 (the model that underpins Chat GPT) could create up to 550 tonnes. Spread across millions of users this may be an acceptable environmental impact, however, it’s worth giving some consideration to.
- Billie-Jay Porter from Vinnies QLD raised a very good point about organisations ensuring that the organisations supplying these tools abide by their supplier obligations. Vinnies QLD has a detailed policy on ethical sourcing which also covers aspects of the Modern Slavery legislation. With news coming into the market from Time and Rolling Stone about the engagement of Kenyan workers to filter some of the most graphic types of content to reduce the potentially toxic outputs that could occur, it does remind us that charities do need to be aware of their supply chains and the ethical sourcing of products.
Certainly in fundraising, I think there is a broad range of potential, from using tools like chat GPT to models that support the richer targeting selections that drive effective fundraising rather than just efficient fundraising, without sacrificing the potential long-term value from supporters. It’s an exciting future ahead, and there are things we can be doing to prepare our organisations for the opportunities ahead. Our recommendations on how to improve your data to take the best advantage of the opportunities that will come:
- Maintain your data well. Clean data is going to be a really big influencer in how effective ML will be identify the signal in the noise.
- Start collecting rich metadata about your campaigns and how you communicated with your donors (did you include a Response Mechanism for these donors, how much were they asked for, what variable paragraph did they receive, what tests did they receive?)
- Collect more dates. Applying date or time stamps to when you collect data makes it much easier to do longitudinal modelling that can look at all the information before our point of ‘success’.