Simple autonomous agents can enhance creative semantic discovery by human groups

The study addresses how simple autonomous agents can affect creative idea discovery within human groups navigating a semantic space. Innovation often requires exploration across many possibilities, and group processes can both mitigate and exacerbate challenges such as groupthink and inefficient diffusion. Collective intelligence benefits from a balance of independence and interdependence; too much interdependence can cause premature convergence on inferior ideas, whereas too little coordination can inhibit exploitation of promising areas. Prior research on social learning has examined network structure, learning strategies, and group size, but has largely overlooked relational properties among candidate ideas, such as semantic similarity and the fact that similar ideas often have similar value and are more easily discovered via incremental improvements. The authors develop a word-based semantic search task to examine how social information and simple bots with interpretable strategies can shape group exploration and exploitation, including under landscapes made more difficult by decoy peaks. They focus on simple, transparent bots using classic NLP (word2vec) to isolate effects on human creativity and interpretability of interventions.

The paper situates its contribution within research on collective intelligence and social learning: (1) the balance of independence and interdependence in groups is critical; excessive social influence can induce groupthink and degrade crowd wisdom, while some interdependence improves coordination and exploitation. (2) Prior work highlights how network structure, social learning strategies, and group size affect collective search and cultural accumulation. More connected networks often aid convergence on easy problems. (3) Existing experimental studies of networked decision-making have underemphasized the role of semantic relationships among ideas. In real settings, similar ideas often have correlated value and can be discovered through incremental changes. (4) The authors focus on simple autonomous agents, extending prior demonstrations that bots can improve coordination and cooperation, while maintaining transparency using word2vec rather than complex LLMs. (5) The study also connects to research using NLP for creativity assessment and to models of rugged fitness landscapes (e.g., NK model), adding controlled decoys to simulate local optima in semantic search. Together, these literatures motivate testing whether simple, local, interpretable bot strategies can enhance group creative search in structured semantic spaces.

Design overview: The authors created a word search game where participants searched a space of 20,000 frequently used English nouns (post-filtering of inappropriate terms) embedded in word2vec (300-dimensional vectors). A single target noun per game had the maximum value (20,000 points), and other nouns received values decreasing monotonically with cosine distance to the target. Eighteen target nouns, chosen to be semantically dispersed and of comparable obscurity, were used across games. Participants’ task was to submit nouns over rounds to achieve high scores. To prevent meta-strategies, displayed scores were multiplied by a random factor in [1,3] per game; analyses use unscaled ranks. Participants and grouping: N = 1,875 participants (Amazon Mechanical Turk) were organized into 125 independent groups of 15 participants each. Each group played five sequential games (25 rounds per game). In each round, participants had 7 seconds to submit a noun; feedback (own and neighbors’ latest noun and score) was shown for 7 seconds. An extra 14 seconds were given in the first round to acclimate. Compensation was $3 base plus up to $11 bonus based on group performance. Data from an additional 25 groups in the tall/wide landscape were collected in a deviation from preregistration but are not included in the 1,875 summary. Networks and social information: Within each game (except solo), participants were embedded in an Erdos–Rényi network (15 human nodes, 20% tie saturation). Participants observed the most recent answers and point values of their immediate neighbors after each round and were always shown the highest-point noun observed so far in their ego-network (including self). In the solo condition, all edges were removed; in the no-bot group condition, only human nodes remained. Incentives were group-based: each participant’s bonus was tied to the maximum group score in each game to discourage producer-scrounger dynamics and redundant repeats. Bot interventions: In three bot conditions, two autonomous agents were added to the 15-human network, yielding 17 nodes. Bots connected to disjoint subsets of humans; edges to bots were not identified to participants. Bots observed their local neighbors’ latest answers and selected a single noun each round to propagate via the other bot, which then immediately broadcast that noun to its own neighbors (one-step relay). Three bot strategies operated solely on human-generated nouns (no target knowledge): (1) Most-similar bot: from the set of neighbors’ nouns, selected the noun with the highest average cosine similarity to the others; (2) Least-similar bot: selected the noun with the lowest average similarity; (3) Random bot: randomly chose one neighbor’s noun. Thus, bots transmitted either consensus-like, outlier-like, or random ideas through the network. Semantic landscapes and decoys: To manipulate difficulty and ruggedness, the authors introduced decoy nouns acting as local optima: the decoy’s neighborhood had artificially boosted ranks, but the target remained globally optimal. Two factors varied between groups: decoy peak height (tall vs short) and width (wide vs narrow), plus a no-decoy condition, producing five landscapes: tall/wide, tall/narrow, short/wide, short/narrow, no-decoy. Wide landscapes boosted many nouns (e.g., 12,000) around the decoy, disrupting the correlation between semantic similarity and value. Each group experienced all five bot conditions within the same landscape type to avoid cross-landscape learning; bot condition order followed a fractional factorial design. Outcome measures: Primary dependent variable was the average cosine similarity of all submitted nouns to the target noun across all 25 rounds in a game. Additional analyses considered the similarity of bot-shared nouns to the target, similarity of participants’ nouns to the decoy, maximum similarity achieved in a game, and behavioral correlations across rounds. Statistical analysis: Bayesian multilevel regressions (RStan/brms) with default priors and standardized dependent variables. Models included fixed effects for bot condition (reference: no-bot or most-similar depending on analysis), landscape factors (reference: short/wide or tall/wide as specified), and interaction terms. Random intercepts for groups and varying intercepts/slopes for target nouns accounted for repeated measures. Posterior means and 95% or 90% highest density intervals (HDIs) were reported. Convergence diagnostics indicated Rhat < 1.05 across models. Network generation used NetworkX. Preregistration (OSF) specified key analyses (notably those underlying Fig. 3 and Fig. 4a).

- Groups outperform individuals working alone in exploring semantic space: Compared to the no-bot group condition, the solo condition had lower average similarity to the target (β_solo = −0.70; 95% HDI [−1.10, −0.33]) in the no-decoy landscape. - Main effects of adding bots (most-similar, least-similar, or random) were not significant overall versus no-bot groups; however, interactions with landscape revealed benefits for the most-similar bot in easier landscapes. - Bot × landscape interactions: The most-similar bot improved group performance in the no-decoy and tall/narrow landscapes (β_Most no decoy = 0.56; 95% HDI [0.05, 1.07]; β_Most tall/narrow = 0.50; 95% HDI [0.00, 1.03]); a similar trend held for short/narrow (β_Most short/narrow = 0.44; 95% HDI [−0.08, 0.97]; 90% HDI [0.00, 0.87]). No other bot–landscape interactions were meaningfully different from zero. - Quality of bot-shared nouns: In easier landscapes (no-decoy, short/narrow, tall/narrow), the most-similar bot propagated nouns with higher similarity to the target compared to the least-similar bot (e.g., β_Least:no decoy = −0.85; 95% HDI [−1.42, −0.25]; β_Least:short/narrow = −0.81; 95% HDI [−1.45, −0.20]; β_Least:tall/narrow = −0.81; 95% HDI [−1.40, −0.20]). The most-similar bot also outperformed the random bot in tall/narrow (β_random:tall/narrow = −0.85; 95% HDI [−1.44, −0.23]) and showed a similar trend in no-decoy (β_random:no decoy = −0.47; 95% HDI [−1.04, 0.14]; 90% HDI [−0.98, 0.01]). - Decoy attraction: The most-similar bot did not increase similarity of participants’ guesses to the decoy noun. In the tall/narrow landscape, least-similar and random bots were more associated with guesses resembling the decoy than the most-similar bot (β_Least:tall/narrow = 0.67; 95% HDI [0.05, 1.31]; β_random:tall/narrow = 0.65; 95% HDI [0.02, 1.26]). - Landscape width disrupts semantic alignment: Participants’ ability to align subsequent guesses with previously observed high-value nouns was reduced in wide landscapes. The correlation between score at round t and cosine similarity of guesses at t and t+1 was lower in wide vs narrow landscapes (β_wide = −0.34; 95% HDI [−0.62, −0.05]). Social information improved this alignment relative to solo (β_solo = −0.33; 95% HDI [−0.52, −0.16] vs no-bot group). No difference between narrow and no-decoy (β_no decoy = −0.06; 95% HDI [−0.45, 0.35]). - Maximum similarity achieved per game did not differ meaningfully across landscapes, but wide landscapes impeded collective exploitation of occasional discoveries. - Individual differences: Performance correlations between solo and social conditions were positive but modest (Pearson r ≈ 0.18–0.23), while correlations among social conditions were stronger (r ≈ 0.38–0.54), suggesting partially distinct traits of solo vs group creativity. Number of unique nouns and stepwise divergence patterns were similar across solo and social settings despite performance differences. - Network position relative to bots did not significantly moderate treatment effects: no 95% HDI differences between participants directly connected to bots vs not connected.

The findings show that simple, transparent autonomous agents can enhance collective creative search in human groups under certain conditions. While groups already outperform isolated individuals in navigating semantic spaces, adding bots that propagate the most semantically central neighbor idea (most-similar bot) further improves performance in easier landscapes (no-decoy and narrow decoy peaks). This suggests that locally amplifying promising, consensus-like ideas improves diffusion and exploitation of high-value regions without herding toward local optima. Conversely, wide decoy landscapes, which artificially boost many nouns and disrupt the correlation between meaning and value, impede groups’ ability to form a coherent mental model and benefit from bot assistance. Social information facilitates semantic alignment across rounds, while solo participants show reduced alignment and exploration. The results indicate that simple bots can leverage humans’ intrinsic problem-solving on navigable landscapes by reducing noise and aiding coordination, without requiring global knowledge of the target. These insights inform how AI interventions might be designed to promote collective intelligence: local, decentralized, and interpretable strategies that reinforce high-value ideas can be beneficial, while environments that scramble signal–value correspondence limit such gains.

This work introduces a controlled semantic search paradigm linking real-world linguistic structure to group innovation and demonstrates that simple, interpretable bots can enhance collective creativity by propagating semantically central, high-value ideas in human networks. The most-similar bot improved average group performance in easier landscapes without steering groups toward decoys, while wide decoys undermined the formation of a usable mental map of the landscape. Contributions include: an experimental platform for group-level semantic navigation; evidence that social information and simple bots aid exploration–exploitation tradeoffs; and characterization of how landscape ruggedness (width) affects collective alignment. Future research should examine: (1) more sophisticated or adaptive bot strategies (including LLMs) and their interpretability; (2) varying network topologies and group sizes to identify context-specific optimal interventions; (3) the interplay between bot-induced connectivity changes and information flow; (4) domain generalization beyond English nouns and to more complex idea spaces; and (5) ethical deployment to avoid reinforcing narrow ideological clusters while enhancing beneficial collaboration in contexts such as citizen science and organizational innovation.

- Bot presence alters network connectivity by adding nodes and edges, confounding pure behavioral effects with structural changes. Effects should be interpreted as supplements to existing networks rather than replacements of members. - Sample size, while substantial, still limits power for some exploratory analyses and raises risk of false discoveries; some interactions significant for average performance were not for best-solution metrics. - The task uses nouns and word2vec embeddings; generalizability to other domains, languages, or more complex ideas is inferred but untested. - Wide decoy manipulation strongly disrupts the value–meaning correlation; ecological validity of such extreme landscapes in real-world settings may vary. - Bots could exacerbate ideological homogeneity in some contexts if amplifying similarity; careful consideration is needed for deployment. - Bots could exchange ideas instantly while humans could not, a design choice that may maximize bot impact but diverges from symmetric human–human information sharing. - Scores presented to participants were scaled by random multipliers per game; although addressed analytically, this adds complexity to participant inference. - Recruitment from MTurk and time constraints (e.g., 7-second rounds) may affect behavior and generalizability.