cs.HCOct 26, 2026
SaveThe Impact of Personalized AI Response Suggestions on Content Creators' Direct Messaging
Organizations: Cornell University, USA
Abstract
Direct messaging is often employed by social media content creators to directly interact with followers. Those creators attempt to engage with followers in private communication, but they often encounter a large volume of messages. Through formative interviews with 15 creators, we found that creators frequently experience fatigue from writing suitable content, as they moderate audience expectations while concurrently integrating their media personas into responses. To explore how AI can address these challenges, we designed a system providing personalized, AI-generated response suggestions tailored to creators' writing styles to maintain their media personas. In a comparison study with 13 creators that examines how the system affects response behaviors, creators valued the system's ability to soften potential impoliteness and anticipated higher response rates, helping them maintain a professional image. However, creators reported additional labor in verifying the authenticity of AI outputs to mitigate risks caused by AI misalignment. Our results indicate that while AI suggestions enhance efficiency, they can undermine a creator's engagement in communication by shifting their attention toward monitoring AI.
Figures & tables
| DM type | Description | Examples |
| Personal Inquiry | • Efforts to establish friendships beyond online interactions (i.e., seeking personal information) or request personal favors to utilize creators’ popularity | • “Can I see more of you?” • “I think short hair suits you.” • “Can you give me contact information?” |
| Creator Persona-related Inquiry | • Efforts to obtain detailed information on the lifestyle and workflow of creators based on their social media posts (Channel-related questions) • Seek advice of message senders, expecting to share a point of view that reflects creators’ lifestyle (Creator lifestyle-related questions) | • “What product do you use to give your lips a glowing appearance?” • “Would you like to recommend brands?” • “Could you provide the recipe for the lemon pie posted yesterday?” |
| Partnership Suggestion | • Suggestion for collaborations in which creators can get benefits from collaborators (e.g., clients, creators) (Benefit-sharing collaborations) • Proposals for asking for free labor for promotion to utilize creators’ popularity (Unpaid promotional partnerships) | • “We’ll send you a product.” • “We love your content, and we would like to send you this.” |
| Compliment | • Positive comment about creators’ social media posting content, lifestyles, and editing styles, showing how message senders appreciate creators’ activities | • “Yo man! You’re super hilarious and your videos never failed to make me laugh. Keep it up” |
| Content Critique | • Suggestions that creators enhance their channels by providing feedback on their posting and editing styles, as well as their choice of attire. | • “Shoutout, would you and Lin ever make a podcast together?” |
| Random | • Sending content (images, anecdotes) the creator does not agree is relevant to them | • Sending videos of themselves playing different violins to music creators |
Table 1. Direct message typology and examples SMCs typically received from audiences, drawn from interviews
| Key Insight | Identified Challenges | Design Goal | Design Components |
| Creators selectively prioritize reactions based on channel goals (§ 3.3.2 ) | Limited time to identify prioritized messages; uncomfortable messages are ignored | DG2, DG4 | Automatic DM classification into six categories; category-filtered inbox and thread summaries (§5.3, §5.3.1) |
| Creators stylize responses for imagined audience groups and channel personas (§ 3.3.3 ) | Difficulty choosing wording and tone according to each audience group, staying in their social media persona without offending them | DG1, DG3 | Persona construction from social media posts; editable persona panel; category-specific writing style settings (§5.2, §5.3) |
| Creators self-limit responses to avoid over-engagement and protect boundaries (§ 3.3.4 ) | Detailed replies invite unwanted follow-ups; adequate-but-guarded responses are effortful | DG3, DG4 | Enhance Message feature; Response Suggestion feature; per-category exclusion of personal data; review panel before sending (§5.3.1, §5.4) |
Table 2. Mapping between formative study insights, design goals, and Creact features
Figure 1. Overall user flow in the Creact interface
Figure 2. The snapshot of the chat conversation board where SMCs can send DMs in Creact
Figure 3. The pipeline to generate two personalized AI-generated response suggestions
| PID | Channel Theme | Gender | # of Followers | Message Replying Frequency |
| U1 | Fashion | M | 16.5k | Once a week |
| U2 | Lifestyle | M | 19k | 2-3 times a week |
| U3 | Lifestyle | M | 958k | 2-3 times a week |
| U4 | Comedy | M | 88.8k | more than 7 times a week |
| U5 | Antique | M | 11.1k | more than 7 times a week |
| U6 | Music | M | 13.1k | 4-6 times a week |
Table 3. Demographic Information of Participants. All participants worked based on Instagram accounts
Figure 4. Main evaluation study procedure
| Creact | Baseline | |||
| DM types | M (SD) | M (SD) | ||
| Personal Inquiries | 146.53 (105.85) | 57.92 (67.61) | 0.0253* | 0.0760 |
| Channel-related questions | 202.23 (148.09) | 100.92 (172.118) | 0.1422 | 0.2844 |
| Creator-lifestyle-related inquiries | 163.23 (82.88) | 65.385 (49.10) | 0.0064* | 0.0514 |
| Benefit-sharing collaboration | 171.23 (59.39) | 101.07 (67.47) | 0.0175** | 0.0755 |
| Unpaid promotional partnership | 118.23 (68.53) | 102.76 (70.71) | 0.6495 | 0.6495 |
Table 4. Comparison of character counts between two conditions across different DM types ( p < 0.01: “**”, p < 0.05: “*”)
| Creact | Baseline | |||
| Component | Item | M (SD) | M (SD) | |
| Social-self confidence | I could present the right language tones for any messages. | 6.385 (0.768) | 5.385 (1.502) | 0.047* |
| I could manage to have every detail of message responses at its best, positive way to match my words. | 5.923 (1.441) | 5.0 (1.633) | 0.037* | |
| Self-image regulation | I could change how I ‘appear’ to audiences when I would need to do so. | 6.385 (0.870) | 4.770 (1.922) | 0.054 |
| Difficulty adapting self-presentation (Reverse): It was difficult to present myself differently according to different message types. | 5.307 (2.213) | 4.846 (1.725) | 0.549 | |
| Social openness | I was not very wary, even though I received unusual messages from somebody I do not know well. | 4.231 (1.964) | 3.770 (.921) | 0.408 |
Table 5. Perceived self-presentation ability items ( Laghi et al., 2011 ) and results by condition
| Prompt Category | Description | Examples |
| Tone Direction | Adjusting language styles or specific topical inclusions based on creator expectations. | “Express gratitude but include no questions”; “Be brief and direct, no emojis.” |
| Output Feedback | Iterative prompting to optimize length or tone; using enhancement features to avoid manual revision. | “Just be brief… I don’t like the last part”; “Fix the overall tone of these sentences.” |
| Linguistic Cues | Specifically requesting the addition of emojis or structural changes like sentence shortening. | “Use scary and smiley emojis”; “Include a love emoji”; “Give me a short reply.” |
| Taking Initiative | Requesting the AI to proactively suggest a more effective or strategically better response. | “Give me a better response” (e.g., for rejecting a partnership invitation). |
Table 6. Types of User Prompts for Refining AI-Generated Responses
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| PID | Channel Theme | Gender | # of Followers | Social Media Platforms Where They Use DM | Message Receiving Frequency | Message Replying Frequency |
| P1 | Fashion | M | 28k | Instagram, TikTok | 4-6 times a week | 4-6 times a week |
| P2 | Cryptocurrency | M | 14k | Instagram, TikTok, X | 4-6 times a week | 4-6 times a week |
| P3 | Comedy | M | 28.8k | Instagram, TikTok | more than 7 times a week | 4-6 times a week |
| P4 | Culture | F | 248k | Instagram, TikTok | more than 7 times a week | Once a week |
| P5 | Fashion | F | 233.7k | TikTok | more than 7 times a week | 2-3 times a week |
| P6 | Shoe-crafting | M | 68.8k | more than 7 times a week | more than 7 times a week |
Table 7. Demographic Information of Formative Study Participants
| Construct | Item Sentence | Scale | (SD) | (SD) |
| Agency | "I wrote the responses" | 1–7 | 6.31 (1.65) | 5.38 (1.70) |
| Control | "I was in control of writing the responses" | 1–7 | 5.46 (1.79) | 5.77 (1.94) |
| Satisfaction | "I’m satisfied with the assistance I received for writing responses from the system" | 1–7 | 5.23 (2.09) | 6.54 (0.52) |
| Future Intent | "I would like to use this system for responding to direct messages in the future" | 1–7 | 5.00 (2.00) | 6.54 (0.66) |
Table 8. Agency, Control, and System Evaluation
| Response behavior patterns | Mean (SD) |
| Use Response Suggestion only | 3.00 (2.83) |
| Use Enhance Message only | 2.15 (2.44) |
| Use Response Suggestion + Enhance Message | 1.31 (1.44) |
| Use Enhance Message + Edit AI-generated content | 0 (0) |
| Use Response Suggestion + Edit AI-generated content | 0.31 (0.63) |
| No AI use | 1.15 (1.14) |
Table 9. The average number of behavior patterns participants showed throughout tasks
| DM types | Direct message stimuli | Responses in Creact | Responses in Baseline |
| Personal Inquiries | do u cry during movies? and if yes WHICH movies? I wont judge I promisee | Hi Frameora, yes I definitely do. I find myself tearing up during emotional films, especially reality shows and love stories (U7). | Yes, I do—many times actually (U5). |
| Channel-related questions | do u have any recs for makeup brands? need ur help pls!! | Hey there! For makeup brands, I totally recommend checking out Fenty Beauty for diversity and quality, and Tarte for fabulous palettes Also, don’t sleep on ColourPop if you’re looking for something affordable! (U13) | No pls (U9). |
| Creator-lifestyle-related inquiries | [Creator name]! What’s ur most used emoji rn and does it actually fit how u feel? | Hey! Right now, I would say my most used emoji has to be the heart. It pretty much fits how I feel about life—grateful and full of love. How about you? What’s your go-to emoji? (U1) | Love emoji! yes it fits my virtue as a loving and kind person (U4) |
| Benefit-sharing collaboration | hey [Creator name], would u be down to try my clothing brand in one of ur vids? | Hey! That sounds interesting. I’d love to check out your clothing brand. Freebies are always a nice touch! Let me know what u have in mind. (U8) | Hello, Yeah i would be down to trying your brand. Sharing my email for more (U12) |
| Unpaid promotional partnership | Hey [Creator name] any chance u’d be up for reviewing my startup’s new creator tool on ur channel? We’d love to offer u affiliate perks if it’s a fit! | Thanks so much for thinking of me and reaching out about your creator tool. I really appreciate the offer, but I’m not taking on affiliate or sponsored reviews at the moment. Wishing you and your team the best with the launch—it sounds like an exciting project! (U6) | Sure, i can do that But i will 1st have to a background check and from there i can decide (U11). |
| Compliment | Saw your new reel [Creator name] and just had to say…the new haircut?? It’s a look, king | Ayy, thank you so much! I’m lovin’ the new look too! Feelin’ fresh and fierce! Ready to take on the world, one event at a time. | Thank you so much! (U12) |
Table 10. Examples of responses written by participants in conditions
More in this topic
Human-AI Collaboration
Human-AI collaboration studies how people and AI systems divide, coordinate, and adapt work to support decisions, creativity, and learning.
See all 442 papers