The measurement behind Original Intelligence
Hupside measures content against what frontier language models would most commonly produce on the same task. Two products apply that measurement to two different questions.
The industry spent this year trying to draw a line between human and machine. Watermarking, flagging, detection. The line doesn't hold, and the watermarks that remain mostly just record that a model was in the room.
We measure something else. Frontier models now define a statistical center for how almost any task gets done. That center is measurable, and so is the distance from it. Everything Hupside builds comes from one innovation: establishing where the models would most commonly land on a given task, and quantifying how far a contribution sits from that point.
One innovation, two applications
Hupchecker → OIQ
HupMapper → Originality Intensity
Measures
A person's performance on a problem-solving challenge
A piece of finished content, up to book length
Baseline
What the major frontier models would most commonly produce for the same challenge
What those models would most commonly produce for the same piece, with no human involvement
Reports
How far beyond Common AI the person scored, plus an OIQ Archetype
Divergence from the Common AI baseline, plus Originality Zones every 400 words
Answers
Can this person solve a problem in original ways that Common AI doesn't already?
How far did this work travel beyond what the models would most commonly produce?
The two products are distinct. They share the measurement layer underneath, which Hupside builds and owns.
Hupchecker and OIQ
A participant completes a structured problem-solving challenge. OIQ reports how far beyond Common AI someone scores.
Divergence means going beyond Common AI, and not every response gets there. A response that sits inside what the models would most commonly produce scores at or near the floor. What OIQ answers is narrow on purpose: can this person solve a problem in original ways that Common AI doesn't already? It is not a measure of general performance, and originality is not always what a task calls for — plenty of work needs to be done exactly the way the models would do it. OIQ speaks to the times when it doesn't.
OIQ is not a norm-referenced test. The comparison is against what current frontier models would most commonly produce, not against a general population, and the reference point moves as the models do.
The OIQ Archetype is an interpretive layer, and we say so. It is an inference drawn from the pattern of a person's OIQ response, mapped onto established findings about how different profiles of idea generation manifest in how people work. That mapping rests on decades of research in creativity and human performance, including our own founding team's contributions to it: the measurement of creative connection across semantic distance, the finding that creative performance is a state as well as a trait — the same person performs differently depending on task, effort and conditions — and the relationship between idea-generation profiles and how people engage with different kinds of work. The archetype is a lens for applying the score. The score is the measurement.
HupMapper and Originality Intensity
HupMapper takes a finished piece of content, up to book length, and measures it against what frontier models would most commonly produce for that same piece with no human involvement at all: no prompting, no iteration, no human writing.
Originality Intensity measures how far the content sits above that baseline. It works like a thermometer. The higher the reading, the more the content diverges from Common AI, whether it was produced with AI or without it.
Originality Zones score the content in 400-word segments, so the reading is granular rather than averaged. Originality is rarely evenly distributed across a long document, and the zone view shows where it actually lives.
A low score is not an accusation. It does not tell you whether AI was used, and it does not judge the work's value. The Gettysburg Address scores 0 — the models can reproduce it, so it sits inside what Common AI already produces. We publish that result because it shows precisely what the score measures and what it doesn't.
On bias — we don't ask AI to judge people
Frontier models carry bias. That's documented and not in dispute. They absorb cultural artifacts from their training data, and those artifacts disadvantage people outside the mainstream of that data.
The problem compounds when those same models are used to evaluate humans. A growing body of work shows that when frontier models are asked to grade, assess, screen or score people and their work, the bias in the model becomes bias in the judgment, falling hardest on underrepresented groups, people with developmental differences, and anyone whose expression sits outside the training distribution. Most AI evaluation products work exactly this way: a frontier model is handed a rubric and asked for a verdict.
Hupside doesn't do that. We never ask a frontier model to evaluate a person or their work. We measure performance and content against what the models most commonly produce — the models are the reference point, not the judge.
The scoring itself is mathematical. No demographic, racial or protected-category information is collected, inferred or available to our models. The score is divergence from the baseline and nothing else.
Three deployments, one closed-loop measurement
Hupside's architecture has two parts. The scoring layer is our proprietary measurement model. It never calls an external LLM API — no frontier model, no third-party endpoint, ever sees submitted content. The platform layer is the SaaS engineering around it: it takes the data, renders the score, and handles account management, permissions and analysis.
Because the scoring layer carries no external dependencies, the two layers can be deployed wherever a customer requires them.
Hosted, for consumer and prosumer. Both layers run on Hupside infrastructure. Individuals and small teams submit work and get a score. Content is processed entirely inside Hupside's environment and is never exposed to an external model. Retention and data handling per Terms of Service and Privacy policy.
Enterprise. A managed installation for a commercial customer, with the data security controls, access management and personal information handling a private-sector enterprise expects.
Secured, for government and high-security environments. Both layers run inside the customer's own secure engineering stack. Scoring happens there. Account management, permissions and analysis happen there. Content never crosses to Hupside infrastructure, and there is no outbound call to anything.
Every configuration is closed loop: submitted content is never exposed to an external model, and scoring never depends on a vendor API. That last point also governs stability. A system built on someone else's model inherits every change they make, silently. Ours doesn't move unless we move it.
Who built it
Dan Johnson is Hupside's Chief Technology Officer and the inventor of the technology behind both products. He co-invented the measurement approach that produces Originality Intelligence Quotient (OIQ) and exclusively invented the measurement approach underlying Originality Intensity.
Dan spent over 10 years building the tooling the field uses for automated measurement of original thinking:
- First author of the paper introducing divergent semantic integration, a method for extracting creativity from narrative text by measuring how much it connects divergent ideas, in Behavior Research Methods
- Co-builder of SemDis and CAP, both open platforms for semantic distance-based and other creativity scoring, free to researchers and used independently of Hupside
- Author of the Maximum Associative Distance method for improving the validity of automated novelty scoring
- Co-author of the field's recommendations for reliable automated assessment of originality, and of the multilingual extension of that work
- Co-principal investigator on two separate three-year and five-year National Science Foundation grants to study creativity in STEM education
Those methods are public. The system Hupside runs is not. Models, calibration, baseline construction, scoring corpus and version control are built and maintained in house.
How we validate
One arm of our validation uses openly available datasets from peer-reviewed research published in top scientific journals. We had no part in the published work, and because the data is open, anyone can replicate our analysis. [Validating Originality Intensity (OI) using an External Validation Approach, 2026]
We also have peer-reviewed published research and additional papers that explain and validate the approach.
Further validation documentation is available to partners under NDA.
Where the research came from
Original Intelligence began as a research collaboration between Dan Johnson and Adam Green, a Georgetown neuroscientist and Hupside co-founder. Adam's work established why distance in idea space matters cognitively. Dan invented the measurement technology that both products run on. The joint papers are cited below.
Development and maintenance of the measurement system sit inside Hupside.
Where the measurement gets applied
Hupside provides a platform for understanding the originality of performance in an AI-infused environment: how far an organization's output extends beyond what the frontier models would most commonly produce.
- Education
- Learning and development
- Change management
- Data governance
- Team building
- Information integrity
Published research
- Johnson et al. — divergent semantic integration, Behavior Research Methods
- Beaty & Johnson — SemDis
- CAP
- Johnson — Maximum Associative Distance
- Beaty, Johnson, Zeitlen & Forthmann — semantic distance and the alternate uses task
- Green — semantic distance and relational cognition
- Moon, Kushlev, … Johnson, … Green — The Link Between Diverse Words and Original Ideas Is Weakening in the AI-Era College Admissions (preprint)
- Beaty, R. E., & Johnson, D. R. (2021). Automating creativity assessment with SemDis: An open platform for computing semantic distance. Behavior Research Methods, 53(2), 757–780. https://doi.org/10.3758/s13428-020-01453-w
- Beaty, R. E., Johnson, D. R., Zeitlen, D. C., & Forthmann, B. (2022). Semantic distanceand the alternate uses task: Recommendations for reliable automated assessment oforiginality. Creativity Research Journal, 34(3),245–260. https://doi.org/10.1080/10400419.2022.2025720
- Green, A. E. (2016). Creativity, within reason: Semantic distance and dynamic statecreativity in relational thinking and reasoning. Current Directions in PsychologicalScience, 25(1), 28–35. https://doi.org/10.1177/0963721415618485
- Johnson, D. R., Kaufman, J. C., Baker, B. S., Patterson, J. D., Barbot, B., Green, A. E.,van Hell, J., Kennedy, E., Sullivan, G. F., Taylor, C. L., Ward, T., & Beaty, R. E. (2023).Divergent semantic integration (DSI): Extracting creativity from narratives withdistributional semantic modeling. Behavior Research Methods, 55(7),3726–3759. https://doi.org/10.3758/s13428-022-01986-2
- Patterson, J. D., Pronchick, J., Panchanadikar, R., Fuge, M., van Hell, J. G., Miller, S.R., Johnson, D. R., & Beaty, R. E. (2025). CAP: The creativity assessment platform foronline testing and automated scoring. Behavior Research Methods, 57(9), Article264. https://doi.org/10.3758/s13428-025-02761-9
- Yu, Y., Beaty, R. E., Forthmann, B., Beeman, M., Cruz, J. H., & Johnson, D. (2025). AMAD method to assess idea novelty: Improving validity of automatic scoring usingmaximum associative distance (MAD). Psychology of Aesthetics, Creativity, and theArts, 19(4), 849–865. https://doi.org/10.1037/aca0000573
- Green & Johnson - The Link Between Diverse Words and Original Ideas Is Weakeningin the AI-Era College Admissions (preprint)
