All patternsTuners
Filters

Filters

Constrain the inputs or the outputs of the AI by source, type, modality, etc

Filters help users constrain and direct the AI upfront. They are a close cousin to parameters, which guide the tokens and focus of the AI. Filters on the other hand provide more strict boundaries, either limiting which references should beconsidered by the AI in the first place, or limiting the output to specific formats or modalities.

Different modes of prompting

Working with AI is a process of tuning inputs and controlling outcomes. There are generally four scenarios users find themselves in:

  • Focusing: When the user has a specific outcome in mind and a clear sense of how to get there (ex: editing an existing piece of content with the intent of changing its tone, format, etc)
  • Navigating: Moments when users have a goal in mind but aren't sure how to get there (ex: when you are searching for the answer to a specific question but don't know which inputs will get you there)
  • Synthesizing: Drawing unknown conclusions or decisions from a fixed set of stuff (ex: looking for trends and insights from a known set of data)
  • Browsing: Pure discovery mode, when you are playing with different inputs and seeing where it takes you (ex: creatively prompting with different tokens to see what images they produce)

Filters can help users in all of these scenarios. When the user has a clear sense of what inputs they want to direct the AI to, they can constrain the data that it pulls in as References. When they know the outcome they are targeting but want to leave the input open, they can constrain the format. Each of these are examples of filters helping the user drive the AI where they want it to go.

Commercial considerations

The commercial implications to this is massive, and it‘s something companies have only just started to explore

Could authors assign meta data to their work that allows it to be licensed to models that follow specific commercial terms, allowing their content to be introduced as a premium filter?

Could this information be traced through the watermarked fingerprints of the sources aggregated in existing LLMs?

Could specific modalities, or modality format (e.g. length) be introduced as a premium filter on the AI outcome? Or perhaps some are only available when using data that attributes the original author.

There are many ways designers could use filters to improve the ethics and the results of the models they are designing, and the interfaces they are designing to interact with those models.

Details and variations

  • Filters are set up-front along with parameters and other specifications for the prompt
  • Combine filters with references to control the type and source of the inputs that the AI incorporates into its result
  • Alternatively, limit the AI's output to certain modalities or audiences
  • Filters can be relaxed on subsequent re-generations just like manual filters on a data set or search

Considerations

Positives

Give users control
Filters allow users to set boundaries around a model to improve the quality and the accuracy of their results. This makes GenerativeAI more useful in a commercial or academic sense, and gives agency to the individuals interacting with the data to maintain human-centeredness

Commercial benefits
The ethical concerns with training data for large models are well documented, but no obvious solution has emerged. Filters might be a viable option. Meta data or even  blockchained relationships between data points could be used to combine or exclude training sources based on permissions, legality, personal data, etc. The GDPR implications alone could be massive. How do we ensure that large swaths of consumers don’t find themselves excluded from this technology because the companies building it didn’t do a sufficiently good job of excluding their personal data? Filters may offer a solution.

Academic benefits
For similar reasons, filters can help control variables in academic contexts, helping to ensure predictable results. Already Perplexity.ai and Julius.ai show how constraining sources to academic papers only can lead to a more rigorous synthesis by the AI.

Potential risks

Risks of narrow datasets
Filters could have the effect of overly constraining the data going into the AI. Google recently had an inadvertent example of this play out in their search results, when users started to test the AI by searching for queries that had very little primary data available (how many rocks to eat daily, for example). As a result, the AI provided ludicrous results, drawn from a very small source of low quality sources.

Use when:
Users have clear constraints that they wish to apply strictly to the AI
Related patterns:
Parameters
Include constraints with your prompt for the AI to reference when generating its result
Follow up
Get more information from the user when the initial prompt isn't sufficiently clear

Examples

Users can add parameters including negative tokens directly into Midjourney's chat interface
Hypotenuse makes it easier to manage common parameters, without the broad flexibility of Midjourney. Presumably their audience is less technical.
Jasper allows users to specify parameters as inputs into their writing
Jasper makes parameters easy to find from the editor as well, so users can specify them before remixing the result
Some writing tools like Jasper give users the option to bias for speed or quality. This would be an interesting pattern to see explored by UI generators, given the different needs for wireframes vs. glossy comps
ReWord lets users add training documents to each prompt as well
Parameters are extended to AutoFill prompts in Coda's interface
No items found.

© Emily Campbell 2024 | Sharable under CC-BY-NC-SA | Contact

Sharable under the Creative Commons CC-BY-NC-SA license