Article

Using AI to help SNAP recipients make sense of “notices"

Table of contents

An early area we’ve been exploring in applying AI to the frictions people face navigating the SNAP (food stamp) program is in helping people make sense of — and act on — notices. Notices can cause confusion and problems for people navigating SNAP, and also can lead to unnecessary call volume and extra work for state agencies.

We’ve prototyped a tool, explored below, that we are sharing with SNAP program experts for initial feedback and, based on what we learn, then plan to test with SNAP recipients.

What are “notices” in SNAP?

Someone interacting with a government program like SNAP faces many interface points. You have:

  • digital interfaces (content web sites, transactional eligibility and enrollment portals)
  • paper interfaces (paper applications, mailed letters)
  • person-interfaces (an agency office lobby, speaking to an eligibility worker on the telephone)
  • physical interfaces (an EBT card itself, the point of sale terminal)

“Notices” are the formal communications that regulations require SNAP agencies to use when doing certain things that affect a person’s SNAP case.

By default, these will be paper notices that come in the mail. But most states also have the option to receive them online via an eligibility portal (specifically, someone receives an email telling them they have a new notice, and then they need to log in to the web site to read it.)

Here’s an example:

SNAP notice of decision for Tennessee

What matters most about notices is they are the formal communication of things like “adverse actions” (say, denying an application, or reducing someone’s benefit amount.)

As such, many requirements apply to the notices because they are in the program design of SNAP how the program makes sure clients are told things when they should be told (timely), are told the right things (adequacy), and can make use of their rights under the program.

In fact, when a SNAP notice fails to meet the requirements for formal communication, this failure alone can be grounds for successfully appealing a negative decision.

Notices are intended to be the artifact that enables SNAP clients to exercise their agency in the program: to understand what is happening, why it is happening, and to formally disagree (by appealing the action, known more specifically in SNAP as requesting a fair hearing.)

Here’s are some examples of events that someone should get a notice about:

  • Denied benefits (including specifically why)
  • Approved for benefits (including for what monthly dollar amount)
  • A request for specific documents (“verifications”)
  • Missing an interview (including how to schedule)
  • Benefit amount reduced (including why)
  • Overpayment (being told you were given too much previously, and so need to pay back)
  • …and more

What are the user problems and opportunities of “notices”? (or, paving paths)

In an area like benefits navigation, you’ll sometimes hear shorthand summaries of problems that are fairly abstract. For SNAP notices, the most common summary is: “notices are confusing to people.”

And that’s true. While there has been much work to improve notices in the past, notices are still often a source of confusion for people. Some reasons why:

  1. SNAP is an intrinsically complicated program, with legal language and special wording often uncommon in everyday life
  2. Most notices have legally required information that can create “cognitive overload” by providing so much information that the most important is hard to suss out (or people may default to stopping reading and calling the agency to ask for help)
  3. Some notices may be completely informational and unlikely to affect benefits, while others are critically important and require immediate action to get or keep benefits

For example, someone posted this notice with the title “Help I’m so confused”:

A reddit post saying Help I'm so confused with a notice of Application approved but $0 benefit amount

The notice says (1) their application was approved, but (2) their household is not eligible and will get $0. It seems reasonable that this person was confused.

The answer? Technically they were under the income limit to qualify—but very close—and when their benefit amount was calculated it was $0 per month. While uncommon, this can happen under the rules.

This notice is a good example of (1) above — that the innately complicated nature of the SNAP can lead to confusing letters.

From general confusion to more specific notice frictions

So “notices are confusing” is a valid statement, at least as a general summary. However, it’s also not as helpful for building useful tools for people as examining more specific instances of problems people have.

In fact, one reason we think a tool helping make sense of SNAP notices is valuable is that we see people on social media (Reddit, Facebook) posting their notices and asking for help understanding them.

With their social media posts, people are already solving this problem, in a bit of a kludgy and difficult way. We're following the signal of that need, and making it easier for many more people, not just the most motivated.

People are already walking this route — we’re paving it for them.

Here are a few examples of more specific notice problems we’ve people have seen:

Confusing informational notices as notices requiring action

A Reddit post with a picture of a notice saying Basic Work Rules

Here’s an example of someone who thought they were losing their SNAP after receiving this notice, so they posted online. Specifically they thought their SNAP would be taken away if they didn’t get a job immediately.

In reality, this notice was informational, and telling them their obligations if they are offered a job. The key takeaway: your benefits are not being changed, but you need to be aware of this if your situation changes.

Confusion about benefit calculation math

Reddit post saying Why has my balance decreased with a picture of a notice showing a reduced benefit amount and calculation

SNAP’s benefit calculation is complicated. We see people post their notices of benefit calculations and try to make sense of whether it is accurate or not.

This usually happens when someone is denied, benefit amounts change, or the amount is less than expected.

This is an especially high-impact notice to help someone with. That’s because if something is wrong in the calculation (for example, if the monthly income is far higher than reality, or if applicable deductions are not counted) then the person has the ability to take action to increase their benefit amount. They can do this by informally requesting a review of the calculation, submitting additional documents proving deductible expenses, or requesting a fair hearing (appealing).


The most relevant information is hidden among less relevant language

Reddit post saying Can someone explain this to me, and showing a notice of expedited SNAP denial

In this example, there are multiple checkbox options with chunks of text next to them that do not apply to this person’s situation. And it states that it is a denial, for expedited SNAP benefits. But if you read closely to where the checkbox is selected, this person will in fact get regular SNAP benefits.

These are just a few illustrations of concrete notice problems people have gone to the Internet for help with. These examples certainly don’t represent all issues, but they offer a view into narrower problems.

The second-order impacts of notice confusion: call center wait times

Confusion caused by notices has direct impacts, like an eligible person not submitting the right documents and being denied.

But it is worth taking a moment to also look at the second-order effects.

From surveying Propel users, we know that the #1 action people take when they have a question about their SNAP benefits is they call the SNAP agency. (In fact, about 50% of people say they call when they have a question.)

On the margin, every instance of confusion coming from notices can lead to:

  • higher call volume to state agencies, which leads to
  • longer wait times for people calling, and
  • less staff capacity available to process applications, renewals, etc. (potentially leading to people having to wait longer to get benefits)

Obviously this is all context setting for…

Our early work applying AI to SNAP notices

We have been actively experimenting with using AI to make sense of SNAP notices, with a focus on reducing confusion, highlighting the right information, and enabling people to take the actions in their best interest (as they see it.)

Notices as a sensible input for large language models

We think SNAP notice confusion is a problem amenable to solutions built on top of this new generation of large language models (LLMs) like ChatGPT, Claude, and Gemini.

That’s because fundamentally most of the problems people have with notices are about language friction, and LLMs have strong capabilities when it comes to processing and manipulating language.

We are testing a variety of models, prompts, designs, external context with real SNAP notices to see what generates helpful output.

Our initial hypothesis, 2 core outputs: “how important is it” and “what do I need to do”

Our working hypothesis for the value of a SNAP notice tool is that it accurately provides two key things for a user:

(1) How important is this notice for me?

Could I lose my benefits if I don’t act immediately? Or, is this general information they think can help me? Or, is this legally required information, but not applicable or important to me?

(2) What action do I need to take (if any)?

Do I need to do something? What? (Make it clear and easy for me.)

These are the two things that, in general, we think apply across effectively all notices.

(A tool successfully extracting these from notices could also form the basis of an even more powerful capability, an agent processing notices in the background and triaging the person’s attention to just those highest-importance notices.)


Example: a notice of case action

Here is one real example of a notice — see if you can quickly make sense of it:

A SNAP notice of case action about missing an interview

A few things are noteworthy that may create friction or confusion in getting the right takeaway point:

  • Roughly the top third of the notice is white space
  • The bottom third of the notice is dense text with nondiscrimination and civil rights complaint information
  • The first 3 paragraphs — and really the 3rd paragraph only — are the most important information
  • The 3rd paragraph has an action (call the phone number) but the number itself is not included, but rather at the top right

When we put this through our prototype AI SNAP notice tool, we get:

A screenshot of a SNAP Notice Helper tool

The output is:

Explanation:

Importance: HIGH importance
This notice is telling you that you missed an interview that's required to keep or start getting SNAP/food assistance benefits. Without completing this interview, your benefits could be denied or stopped.

What you need to do:
• Call (850) 300-4323 right away to reschedule your interview
• This is the only thing you need to do, but it's very important to do it soon

I know keeping track of appointments can be challenging when you have a lot going on. The good news is that fixing this is pretty simple - just one phone call to reschedule. Try to make this call as soon as you can to avoid any problems with your benefits.


Some technical details on the above as part of sharing how we’re building:

  • The model used here was Anthropic’s Claude 3.5 Sonnet (newest)
  • The prompt used here is the following:
    • “You are a legal aid attorney specializing in SNAP benefits. You are given a SNAP notice and an optional question about it. Your response should be short and include two seconts. 1. How important is this notice? (related to the user getting or keeping benefits). Say high importance, medium, or low. And say why. And 2. What the user needs should do, if anything. When you give this action make it as clear and easy as possible for the person to act. Other details: speak in an empathetic tone, using 6th grade reading level language. Generate bullets if that is the most effective way to convey the information. Lastly, remember that you're speaking with someone who has a lot to manage in their life, and you want to inform them without overwhelming them.”
  • This is prototyped using Streamlit, an open source tool enabling rapid iteration on applications using AI models

We are running similar tests for output across different models, prompts, and examples of real SNAP notices we have or find.

Enabling people to ask specific questions about a notice

Separately we also think it’s important for people to be able to ask specific questions about a notice.

That’s both because we see people come with specific questions, and also because some questions may not be answerable by the information in the notice itself. A question may therefore entail an off-ramp to another option for help.

Here’s another example of a notice running through the tool, but this time with a real question from a user who posted this on social media:

Question: I just got this letter from SNAP. I normally get my snap at the end of the month. Does this mean effective immediately or do I still get June and July's?

A SNAP notice about a case showing they are not eligible

Our prototype tool:

A screenshot of a SNAP Notice Helper tool

The output:

Explanation:

This notice is important regarding your SNAP food benefits. Based on the information provided:

• The notice states you are not eligible for SNAP benefits going forward, citing certain Texas policy codes.
• However, it appears to indicate you will continue receiving benefits for a certain number of months due to your dependent children.

So to directly answer your question:

• You should still receive your SNAP benefits for June and July as scheduled.
• But after that period, your SNAP benefits will likely end unless you take further action.

My suggestion would be:
• Read the full notice carefully to understand the specific reason given for terminating your benefits after the designated months.
• Contact your caseworker or the number provided on the notice to get clarity on next steps if you believe you should still qualify.
• Gather any documentation that may support your continued eligibility for SNAP.
• You likely have the option to appeal this decision if you disagree with the determination.

The most important things are to not panic, make sure you receive the benefits for the months allowed, and promptly look into appealing or reapplying if you feel this termination is incorrect based on your current circumstances.

In both of these cases, what the model provides appears to be more helpful to a person than what they can get from simply reading the notice.

We are excited by early successes we see like this, using real notices and real examples of questions.

Safety and risk considerations

With any application of AI, it’s important to assess safety and risk considerations up front. Obviously safety is of higher concern when the potential user base (like SNAP users) would face more extreme negative consequences from failures or problems of a tool.

Some safety and risk elements we are assessing:

1. We are not deploying to users in this state. Right now we are using a combination of real examples we identify (e.g. from social media) as well as feedback from experts in SNAP to harden this tool and proactively identify risks before moving forward.

2. Notices may provide a safer deployment path for AI than other interfaces. Because most of the use of AI here is on processing information already included on the notice itself, this may be safer than other deployment paths, such as a chatbot generating novel answers to open-ended questions. Hallucination risk is lower by definition due to this. And, if needed, we can further enforce this approach in the prompt or by techniques like chaining (e.g. we could have a separate model call evaluate the response to ensure all provided information is found on the original notice in some form.)

3. Risk of missing important information by filtering information. Information one person might find extraneous may be critically valuable to another. For example, required boilerplate information may not be commonly utilized, but may be of extremely high importance in certain circumstances. (The corollary risk is that by overindexing to providing lots of information, some number of people may be overloaded, a phenomenon we see in questions today.) Mitigation strategies we might consider include: always reinforcing that a user should read the entire notice; identifying less common edge cases and including a proactive interaction for people to get more information in those cases (e.g. “this notice also gives info about X, Y, and Z — want to hear more about any of these?”) We want to work on finding the right balance with feedback from program experts and users.

4. PII and sensitive info considerations. By using external model APIs with notices, we need to identify strategies to either redact or ensure no persistence of personally identifying information (PII) with those external providers. Running a local redaction/deidentification model (such as Microsoft’s open source Presidio) to do that is one strategy. We also note that many users already redact information using drawing tools when they post information online.

5. Incorrect notices. If a notice contains incorrect information itself, and our tool is predicated on conveying information as it is on the notice, this could lead to adverse outcomes for users. It is worth noting that, absent a tool, this situation is already an adverse outcome for users. As such, to mitigate this risk, we are considering whether we can include additional information or run additional checks in highly consequential situations that could inform the user if the notice itself appears erroneous or in violation of policy.

Other things we’re thinking about

A few other things to mention that we are thinking about:

1. Bringing in external information in-context with the notice. If a notice says to call a phone number to reschedule an interview, but that phone number consistently deflects calls (“sorry, call volume is too high right now, try again later”) then meeting the user’s need involves bringing in additional information and context. This would be more valuable, but also presents some risks relative to strictly narrowing output to what is on the notice.

2. A notice tool may complement navigation tools more generally (or, “have you checked your notices?” is a common navigation question) While this presumes a person is coming with a question about a notice, it may provide a foundation for broader SNAP navigation assistance, given that many problems’ first step is assessing any notices received recently. (For example, one of the most common questions: “why didn’t I get a deposit this month?”)

3. An ideal notice experience may be entirely passive in the background. While we can enable people to easily take pictures of notices they receive on their phone (or screenshots from their portal) and share those, there exists a better ideal, much more involved, where an AI tool provides background support passively reading notices and alerting the person for action, rather than requiring the person to proactively seek out and use a tool each and every time. This is certainly not in our short term scope, but it is worth keeping in mind what the actual best possible experience could be.

Feedback welcomed!

We welcome feedback, particularly from two groups:

1. People who provide help with SNAP notices: If you are someone who helps SNAP clients with their notices (a public benefits/legal aid attorney, a SNAP outreach worker, even eligibility workers) we’d love to hear from you! We can also provide access to our alpha tool so you can put in notices you may have and test for scenarios you have.

2. AI researchers/practitioners: If you have ideas on how we can more effectively make use of capabilities and evaluate and ensure safety, please reach out!

You can connect with us by emailing me (Dave) at dave.guarino@joinpropel.com.

References and further reading